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) 2000-2008, Intel Corporation, all rights reserved.
14 // Copyright (C) 2009, Willow Garage Inc., all rights reserved.
15 // Copyright (C) 2014-2015, Itseez Inc., all rights reserved.
16 // Third party copyrights are property of their respective owners.
17 //
18 // Redistribution and use in source and binary forms, with or without modification,
19 // are permitted provided that the following conditions are met:
20 //
21 //   * Redistribution's of source code must retain the above copyright notice,
22 //     this list of conditions and the following disclaimer.
23 //
24 //   * Redistribution's in binary form must reproduce the above copyright notice,
25 //     this list of conditions and the following disclaimer in the documentation
26 //     and/or other materials provided with the distribution.
27 //
28 //   * The name of the copyright holders may not be used to endorse or promote products
29 //     derived from this software without specific prior written permission.
30 //
31 // This software is provided by the copyright holders and contributors "as is" and
32 // any express or implied warranties, including, but not limited to, the implied
33 // warranties of merchantability and fitness for a particular purpose are disclaimed.
34 // In no event shall the Intel Corporation or contributors be liable for any direct,
35 // indirect, incidental, special, exemplary, or consequential damages
36 // (including, but not limited to, procurement of substitute goods or services;
37 // loss of use, data, or profits; or business interruption) however caused
38 // and on any theory of liability, whether in contract, strict liability,
39 // or tort (including negligence or otherwise) arising in any way out of
40 // the use of this software, even if advised of the possibility of such damage.
41 //
42 //M*/
43 
44 #include "precomp.hpp"
45 #include "opencl_kernels_imgproc.hpp"
46 
47 /*
48  * This file includes the code, contributed by Simon Perreault
49  * (the function icvMedianBlur_8u_O1)
50  *
51  * Constant-time median filtering -- http://nomis80.org/ctmf.html
52  * Copyright (C) 2006 Simon Perreault
53  *
54  * Contact:
55  *  Laboratoire de vision et systemes numeriques
56  *  Pavillon Adrien-Pouliot
57  *  Universite Laval
58  *  Sainte-Foy, Quebec, Canada
59  *  G1K 7P4
60  *
61  *  perreaul@gel.ulaval.ca
62  */
63 
64 namespace cv
65 {
66 
67 /****************************************************************************************\
68                                          Box Filter
69 \****************************************************************************************/
70 
71 template<typename T, typename ST>
72 struct RowSum :
73         public BaseRowFilter
74 {
RowSumcv::RowSum75     RowSum( int _ksize, int _anchor ) :
76         BaseRowFilter()
77     {
78         ksize = _ksize;
79         anchor = _anchor;
80     }
81 
operator ()cv::RowSum82     virtual void operator()(const uchar* src, uchar* dst, int width, int cn)
83     {
84         const T* S = (const T*)src;
85         ST* D = (ST*)dst;
86         int i = 0, k, ksz_cn = ksize*cn;
87 
88         width = (width - 1)*cn;
89         for( k = 0; k < cn; k++, S++, D++ )
90         {
91             ST s = 0;
92             for( i = 0; i < ksz_cn; i += cn )
93                 s += S[i];
94             D[0] = s;
95             for( i = 0; i < width; i += cn )
96             {
97                 s += S[i + ksz_cn] - S[i];
98                 D[i+cn] = s;
99             }
100         }
101     }
102 };
103 
104 
105 template<typename ST, typename T>
106 struct ColumnSum :
107         public BaseColumnFilter
108 {
ColumnSumcv::ColumnSum109     ColumnSum( int _ksize, int _anchor, double _scale ) :
110         BaseColumnFilter()
111     {
112         ksize = _ksize;
113         anchor = _anchor;
114         scale = _scale;
115         sumCount = 0;
116     }
117 
resetcv::ColumnSum118     virtual void reset() { sumCount = 0; }
119 
operator ()cv::ColumnSum120     virtual void operator()(const uchar** src, uchar* dst, int dststep, int count, int width)
121     {
122         int i;
123         ST* SUM;
124         bool haveScale = scale != 1;
125         double _scale = scale;
126 
127         if( width != (int)sum.size() )
128         {
129             sum.resize(width);
130             sumCount = 0;
131         }
132 
133         SUM = &sum[0];
134         if( sumCount == 0 )
135         {
136             memset((void*)SUM, 0, width*sizeof(ST));
137 
138             for( ; sumCount < ksize - 1; sumCount++, src++ )
139             {
140                 const ST* Sp = (const ST*)src[0];
141                 for( i = 0; i <= width - 2; i += 2 )
142                 {
143                     ST s0 = SUM[i] + Sp[i], s1 = SUM[i+1] + Sp[i+1];
144                     SUM[i] = s0; SUM[i+1] = s1;
145                 }
146 
147                 for( ; i < width; i++ )
148                     SUM[i] += Sp[i];
149             }
150         }
151         else
152         {
153             CV_Assert( sumCount == ksize-1 );
154             src += ksize-1;
155         }
156 
157         for( ; count--; src++ )
158         {
159             const ST* Sp = (const ST*)src[0];
160             const ST* Sm = (const ST*)src[1-ksize];
161             T* D = (T*)dst;
162             if( haveScale )
163             {
164                 for( i = 0; i <= width - 2; i += 2 )
165                 {
166                     ST s0 = SUM[i] + Sp[i], s1 = SUM[i+1] + Sp[i+1];
167                     D[i] = saturate_cast<T>(s0*_scale);
168                     D[i+1] = saturate_cast<T>(s1*_scale);
169                     s0 -= Sm[i]; s1 -= Sm[i+1];
170                     SUM[i] = s0; SUM[i+1] = s1;
171                 }
172 
173                 for( ; i < width; i++ )
174                 {
175                     ST s0 = SUM[i] + Sp[i];
176                     D[i] = saturate_cast<T>(s0*_scale);
177                     SUM[i] = s0 - Sm[i];
178                 }
179             }
180             else
181             {
182                 for( i = 0; i <= width - 2; i += 2 )
183                 {
184                     ST s0 = SUM[i] + Sp[i], s1 = SUM[i+1] + Sp[i+1];
185                     D[i] = saturate_cast<T>(s0);
186                     D[i+1] = saturate_cast<T>(s1);
187                     s0 -= Sm[i]; s1 -= Sm[i+1];
188                     SUM[i] = s0; SUM[i+1] = s1;
189                 }
190 
191                 for( ; i < width; i++ )
192                 {
193                     ST s0 = SUM[i] + Sp[i];
194                     D[i] = saturate_cast<T>(s0);
195                     SUM[i] = s0 - Sm[i];
196                 }
197             }
198             dst += dststep;
199         }
200     }
201 
202     double scale;
203     int sumCount;
204     std::vector<ST> sum;
205 };
206 
207 
208 template<>
209 struct ColumnSum<int, uchar> :
210         public BaseColumnFilter
211 {
ColumnSumcv::ColumnSum212     ColumnSum( int _ksize, int _anchor, double _scale ) :
213         BaseColumnFilter()
214     {
215         ksize = _ksize;
216         anchor = _anchor;
217         scale = _scale;
218         sumCount = 0;
219     }
220 
resetcv::ColumnSum221     virtual void reset() { sumCount = 0; }
222 
operator ()cv::ColumnSum223     virtual void operator()(const uchar** src, uchar* dst, int dststep, int count, int width)
224     {
225         int i;
226         int* SUM;
227         bool haveScale = scale != 1;
228         double _scale = scale;
229 
230         #if CV_SSE2
231             bool haveSSE2 = checkHardwareSupport(CV_CPU_SSE2);
232         #endif
233 
234         if( width != (int)sum.size() )
235         {
236             sum.resize(width);
237             sumCount = 0;
238         }
239 
240         SUM = &sum[0];
241         if( sumCount == 0 )
242         {
243             memset((void*)SUM, 0, width*sizeof(int));
244             for( ; sumCount < ksize - 1; sumCount++, src++ )
245             {
246                 const int* Sp = (const int*)src[0];
247                 i = 0;
248                 #if CV_SSE2
249                 if(haveSSE2)
250                 {
251                     for( ; i <= width-4; i+=4 )
252                     {
253                         __m128i _sum = _mm_loadu_si128((const __m128i*)(SUM+i));
254                         __m128i _sp = _mm_loadu_si128((const __m128i*)(Sp+i));
255                         _mm_storeu_si128((__m128i*)(SUM+i),_mm_add_epi32(_sum, _sp));
256                     }
257                 }
258                 #elif CV_NEON
259                 for( ; i <= width - 4; i+=4 )
260                     vst1q_s32(SUM + i, vaddq_s32(vld1q_s32(SUM + i), vld1q_s32(Sp + i)));
261                 #endif
262                 for( ; i < width; i++ )
263                     SUM[i] += Sp[i];
264             }
265         }
266         else
267         {
268             CV_Assert( sumCount == ksize-1 );
269             src += ksize-1;
270         }
271 
272         for( ; count--; src++ )
273         {
274             const int* Sp = (const int*)src[0];
275             const int* Sm = (const int*)src[1-ksize];
276             uchar* D = (uchar*)dst;
277             if( haveScale )
278             {
279                 i = 0;
280                 #if CV_SSE2
281                 if(haveSSE2)
282                 {
283                     const __m128 scale4 = _mm_set1_ps((float)_scale);
284                     for( ; i <= width-8; i+=8 )
285                     {
286                         __m128i _sm  = _mm_loadu_si128((const __m128i*)(Sm+i));
287                         __m128i _sm1  = _mm_loadu_si128((const __m128i*)(Sm+i+4));
288 
289                         __m128i _s0  = _mm_add_epi32(_mm_loadu_si128((const __m128i*)(SUM+i)),
290                                                      _mm_loadu_si128((const __m128i*)(Sp+i)));
291                         __m128i _s01  = _mm_add_epi32(_mm_loadu_si128((const __m128i*)(SUM+i+4)),
292                                                       _mm_loadu_si128((const __m128i*)(Sp+i+4)));
293 
294                         __m128i _s0T = _mm_cvtps_epi32(_mm_mul_ps(scale4, _mm_cvtepi32_ps(_s0)));
295                         __m128i _s0T1 = _mm_cvtps_epi32(_mm_mul_ps(scale4, _mm_cvtepi32_ps(_s01)));
296 
297                         _s0T = _mm_packs_epi32(_s0T, _s0T1);
298 
299                         _mm_storel_epi64((__m128i*)(D+i), _mm_packus_epi16(_s0T, _s0T));
300 
301                         _mm_storeu_si128((__m128i*)(SUM+i), _mm_sub_epi32(_s0,_sm));
302                         _mm_storeu_si128((__m128i*)(SUM+i+4),_mm_sub_epi32(_s01,_sm1));
303                     }
304                 }
305                 #elif CV_NEON
306                 float32x4_t v_scale = vdupq_n_f32((float)_scale);
307                 for( ; i <= width-8; i+=8 )
308                 {
309                     int32x4_t v_s0 = vaddq_s32(vld1q_s32(SUM + i), vld1q_s32(Sp + i));
310                     int32x4_t v_s01 = vaddq_s32(vld1q_s32(SUM + i + 4), vld1q_s32(Sp + i + 4));
311 
312                     uint32x4_t v_s0d = cv_vrndq_u32_f32(vmulq_f32(vcvtq_f32_s32(v_s0), v_scale));
313                     uint32x4_t v_s01d = cv_vrndq_u32_f32(vmulq_f32(vcvtq_f32_s32(v_s01), v_scale));
314 
315                     uint16x8_t v_dst = vcombine_u16(vqmovn_u32(v_s0d), vqmovn_u32(v_s01d));
316                     vst1_u8(D + i, vqmovn_u16(v_dst));
317 
318                     vst1q_s32(SUM + i, vsubq_s32(v_s0, vld1q_s32(Sm + i)));
319                     vst1q_s32(SUM + i + 4, vsubq_s32(v_s01, vld1q_s32(Sm + i + 4)));
320                 }
321                 #endif
322                 for( ; i < width; i++ )
323                 {
324                     int s0 = SUM[i] + Sp[i];
325                     D[i] = saturate_cast<uchar>(s0*_scale);
326                     SUM[i] = s0 - Sm[i];
327                 }
328             }
329             else
330             {
331                 i = 0;
332                 #if CV_SSE2
333                 if(haveSSE2)
334                 {
335                     for( ; i <= width-8; i+=8 )
336                     {
337                         __m128i _sm  = _mm_loadu_si128((const __m128i*)(Sm+i));
338                         __m128i _sm1  = _mm_loadu_si128((const __m128i*)(Sm+i+4));
339 
340                         __m128i _s0  = _mm_add_epi32(_mm_loadu_si128((const __m128i*)(SUM+i)),
341                                                      _mm_loadu_si128((const __m128i*)(Sp+i)));
342                         __m128i _s01  = _mm_add_epi32(_mm_loadu_si128((const __m128i*)(SUM+i+4)),
343                                                       _mm_loadu_si128((const __m128i*)(Sp+i+4)));
344 
345                         __m128i _s0T = _mm_packs_epi32(_s0, _s01);
346 
347                         _mm_storel_epi64((__m128i*)(D+i), _mm_packus_epi16(_s0T, _s0T));
348 
349                         _mm_storeu_si128((__m128i*)(SUM+i), _mm_sub_epi32(_s0,_sm));
350                         _mm_storeu_si128((__m128i*)(SUM+i+4),_mm_sub_epi32(_s01,_sm1));
351                     }
352                 }
353                 #elif CV_NEON
354                 for( ; i <= width-8; i+=8 )
355                 {
356                     int32x4_t v_s0 = vaddq_s32(vld1q_s32(SUM + i), vld1q_s32(Sp + i));
357                     int32x4_t v_s01 = vaddq_s32(vld1q_s32(SUM + i + 4), vld1q_s32(Sp + i + 4));
358 
359                     uint16x8_t v_dst = vcombine_u16(vqmovun_s32(v_s0), vqmovun_s32(v_s01));
360                     vst1_u8(D + i, vqmovn_u16(v_dst));
361 
362                     vst1q_s32(SUM + i, vsubq_s32(v_s0, vld1q_s32(Sm + i)));
363                     vst1q_s32(SUM + i + 4, vsubq_s32(v_s01, vld1q_s32(Sm + i + 4)));
364                 }
365                 #endif
366 
367                 for( ; i < width; i++ )
368                 {
369                     int s0 = SUM[i] + Sp[i];
370                     D[i] = saturate_cast<uchar>(s0);
371                     SUM[i] = s0 - Sm[i];
372                 }
373             }
374             dst += dststep;
375         }
376     }
377 
378     double scale;
379     int sumCount;
380     std::vector<int> sum;
381 };
382 
383 template<>
384 struct ColumnSum<int, short> :
385         public BaseColumnFilter
386 {
ColumnSumcv::ColumnSum387     ColumnSum( int _ksize, int _anchor, double _scale ) :
388         BaseColumnFilter()
389     {
390         ksize = _ksize;
391         anchor = _anchor;
392         scale = _scale;
393         sumCount = 0;
394     }
395 
resetcv::ColumnSum396     virtual void reset() { sumCount = 0; }
397 
operator ()cv::ColumnSum398     virtual void operator()(const uchar** src, uchar* dst, int dststep, int count, int width)
399     {
400         int i;
401         int* SUM;
402         bool haveScale = scale != 1;
403         double _scale = scale;
404 
405         #if CV_SSE2
406             bool haveSSE2 = checkHardwareSupport(CV_CPU_SSE2);
407         #endif
408 
409         if( width != (int)sum.size() )
410         {
411             sum.resize(width);
412             sumCount = 0;
413         }
414         SUM = &sum[0];
415         if( sumCount == 0 )
416         {
417             memset((void*)SUM, 0, width*sizeof(int));
418             for( ; sumCount < ksize - 1; sumCount++, src++ )
419             {
420                 const int* Sp = (const int*)src[0];
421                 i = 0;
422                 #if CV_SSE2
423                 if(haveSSE2)
424                 {
425                     for( ; i <= width-4; i+=4 )
426                     {
427                         __m128i _sum = _mm_loadu_si128((const __m128i*)(SUM+i));
428                         __m128i _sp = _mm_loadu_si128((const __m128i*)(Sp+i));
429                         _mm_storeu_si128((__m128i*)(SUM+i),_mm_add_epi32(_sum, _sp));
430                     }
431                 }
432                 #elif CV_NEON
433                 for( ; i <= width - 4; i+=4 )
434                     vst1q_s32(SUM + i, vaddq_s32(vld1q_s32(SUM + i), vld1q_s32(Sp + i)));
435                 #endif
436                 for( ; i < width; i++ )
437                     SUM[i] += Sp[i];
438             }
439         }
440         else
441         {
442             CV_Assert( sumCount == ksize-1 );
443             src += ksize-1;
444         }
445 
446         for( ; count--; src++ )
447         {
448             const int* Sp = (const int*)src[0];
449             const int* Sm = (const int*)src[1-ksize];
450             short* D = (short*)dst;
451             if( haveScale )
452             {
453                 i = 0;
454                 #if CV_SSE2
455                 if(haveSSE2)
456                 {
457                     const __m128 scale4 = _mm_set1_ps((float)_scale);
458                     for( ; i <= width-8; i+=8 )
459                     {
460                         __m128i _sm   = _mm_loadu_si128((const __m128i*)(Sm+i));
461                         __m128i _sm1  = _mm_loadu_si128((const __m128i*)(Sm+i+4));
462 
463                         __m128i _s0  = _mm_add_epi32(_mm_loadu_si128((const __m128i*)(SUM+i)),
464                                                      _mm_loadu_si128((const __m128i*)(Sp+i)));
465                         __m128i _s01  = _mm_add_epi32(_mm_loadu_si128((const __m128i*)(SUM+i+4)),
466                                                       _mm_loadu_si128((const __m128i*)(Sp+i+4)));
467 
468                         __m128i _s0T  = _mm_cvtps_epi32(_mm_mul_ps(scale4, _mm_cvtepi32_ps(_s0)));
469                         __m128i _s0T1 = _mm_cvtps_epi32(_mm_mul_ps(scale4, _mm_cvtepi32_ps(_s01)));
470 
471                         _mm_storeu_si128((__m128i*)(D+i), _mm_packs_epi32(_s0T, _s0T1));
472 
473                         _mm_storeu_si128((__m128i*)(SUM+i),_mm_sub_epi32(_s0,_sm));
474                         _mm_storeu_si128((__m128i*)(SUM+i+4), _mm_sub_epi32(_s01,_sm1));
475                     }
476                 }
477                 #elif CV_NEON
478                 float32x4_t v_scale = vdupq_n_f32((float)_scale);
479                 for( ; i <= width-8; i+=8 )
480                 {
481                     int32x4_t v_s0 = vaddq_s32(vld1q_s32(SUM + i), vld1q_s32(Sp + i));
482                     int32x4_t v_s01 = vaddq_s32(vld1q_s32(SUM + i + 4), vld1q_s32(Sp + i + 4));
483 
484                     int32x4_t v_s0d = cv_vrndq_s32_f32(vmulq_f32(vcvtq_f32_s32(v_s0), v_scale));
485                     int32x4_t v_s01d = cv_vrndq_s32_f32(vmulq_f32(vcvtq_f32_s32(v_s01), v_scale));
486                     vst1q_s16(D + i, vcombine_s16(vqmovn_s32(v_s0d), vqmovn_s32(v_s01d)));
487 
488                     vst1q_s32(SUM + i, vsubq_s32(v_s0, vld1q_s32(Sm + i)));
489                     vst1q_s32(SUM + i + 4, vsubq_s32(v_s01, vld1q_s32(Sm + i + 4)));
490                 }
491                 #endif
492                 for( ; i < width; i++ )
493                 {
494                     int s0 = SUM[i] + Sp[i];
495                     D[i] = saturate_cast<short>(s0*_scale);
496                     SUM[i] = s0 - Sm[i];
497                 }
498             }
499             else
500             {
501                 i = 0;
502                 #if CV_SSE2
503                 if(haveSSE2)
504                 {
505                     for( ; i <= width-8; i+=8 )
506                     {
507 
508                         __m128i _sm  = _mm_loadu_si128((const __m128i*)(Sm+i));
509                         __m128i _sm1  = _mm_loadu_si128((const __m128i*)(Sm+i+4));
510 
511                         __m128i _s0  = _mm_add_epi32(_mm_loadu_si128((const __m128i*)(SUM+i)),
512                                                      _mm_loadu_si128((const __m128i*)(Sp+i)));
513                         __m128i _s01  = _mm_add_epi32(_mm_loadu_si128((const __m128i*)(SUM+i+4)),
514                                                       _mm_loadu_si128((const __m128i*)(Sp+i+4)));
515 
516                         _mm_storeu_si128((__m128i*)(D+i), _mm_packs_epi32(_s0, _s01));
517 
518                         _mm_storeu_si128((__m128i*)(SUM+i), _mm_sub_epi32(_s0,_sm));
519                         _mm_storeu_si128((__m128i*)(SUM+i+4),_mm_sub_epi32(_s01,_sm1));
520                     }
521                 }
522                 #elif CV_NEON
523                 for( ; i <= width-8; i+=8 )
524                 {
525                     int32x4_t v_s0 = vaddq_s32(vld1q_s32(SUM + i), vld1q_s32(Sp + i));
526                     int32x4_t v_s01 = vaddq_s32(vld1q_s32(SUM + i + 4), vld1q_s32(Sp + i + 4));
527 
528                     vst1q_s16(D + i, vcombine_s16(vqmovn_s32(v_s0), vqmovn_s32(v_s01)));
529 
530                     vst1q_s32(SUM + i, vsubq_s32(v_s0, vld1q_s32(Sm + i)));
531                     vst1q_s32(SUM + i + 4, vsubq_s32(v_s01, vld1q_s32(Sm + i + 4)));
532                 }
533                 #endif
534 
535                 for( ; i < width; i++ )
536                 {
537                     int s0 = SUM[i] + Sp[i];
538                     D[i] = saturate_cast<short>(s0);
539                     SUM[i] = s0 - Sm[i];
540                 }
541             }
542             dst += dststep;
543         }
544     }
545 
546     double scale;
547     int sumCount;
548     std::vector<int> sum;
549 };
550 
551 
552 template<>
553 struct ColumnSum<int, ushort> :
554         public BaseColumnFilter
555 {
ColumnSumcv::ColumnSum556     ColumnSum( int _ksize, int _anchor, double _scale ) :
557         BaseColumnFilter()
558     {
559         ksize = _ksize;
560         anchor = _anchor;
561         scale = _scale;
562         sumCount = 0;
563     }
564 
resetcv::ColumnSum565     virtual void reset() { sumCount = 0; }
566 
operator ()cv::ColumnSum567     virtual void operator()(const uchar** src, uchar* dst, int dststep, int count, int width)
568     {
569         int i;
570         int* SUM;
571         bool haveScale = scale != 1;
572         double _scale = scale;
573         #if CV_SSE2
574                 bool haveSSE2 =  checkHardwareSupport(CV_CPU_SSE2);
575         #endif
576 
577         if( width != (int)sum.size() )
578         {
579             sum.resize(width);
580             sumCount = 0;
581         }
582         SUM = &sum[0];
583         if( sumCount == 0 )
584         {
585             memset((void*)SUM, 0, width*sizeof(int));
586             for( ; sumCount < ksize - 1; sumCount++, src++ )
587             {
588                 const int* Sp = (const int*)src[0];
589                 i = 0;
590                 #if CV_SSE2
591                 if(haveSSE2)
592                 {
593                     for( ; i < width-4; i+=4 )
594                     {
595                         __m128i _sum = _mm_loadu_si128((const __m128i*)(SUM+i));
596                         __m128i _sp = _mm_loadu_si128((const __m128i*)(Sp+i));
597                         _mm_storeu_si128((__m128i*)(SUM+i), _mm_add_epi32(_sum, _sp));
598                     }
599                 }
600                 #elif CV_NEON
601                 for( ; i <= width - 4; i+=4 )
602                     vst1q_s32(SUM + i, vaddq_s32(vld1q_s32(SUM + i), vld1q_s32(Sp + i)));
603                 #endif
604                 for( ; i < width; i++ )
605                     SUM[i] += Sp[i];
606             }
607         }
608         else
609         {
610             CV_Assert( sumCount == ksize-1 );
611             src += ksize-1;
612         }
613 
614         for( ; count--; src++ )
615         {
616             const int* Sp = (const int*)src[0];
617             const int* Sm = (const int*)src[1-ksize];
618             ushort* D = (ushort*)dst;
619             if( haveScale )
620             {
621                 i = 0;
622                 #if CV_SSE2
623                 if(haveSSE2)
624                 {
625                     const __m128 scale4 = _mm_set1_ps((float)_scale);
626                     const __m128i delta0 = _mm_set1_epi32(0x8000);
627                     const __m128i delta1 = _mm_set1_epi32(0x80008000);
628 
629                     for( ; i < width-4; i+=4)
630                     {
631                         __m128i _sm   = _mm_loadu_si128((const __m128i*)(Sm+i));
632                         __m128i _s0   = _mm_add_epi32(_mm_loadu_si128((const __m128i*)(SUM+i)),
633                                                       _mm_loadu_si128((const __m128i*)(Sp+i)));
634 
635                         __m128i _res = _mm_cvtps_epi32(_mm_mul_ps(scale4, _mm_cvtepi32_ps(_s0)));
636 
637                         _res = _mm_sub_epi32(_res, delta0);
638                         _res = _mm_add_epi16(_mm_packs_epi32(_res, _res), delta1);
639 
640                         _mm_storel_epi64((__m128i*)(D+i), _res);
641                         _mm_storeu_si128((__m128i*)(SUM+i), _mm_sub_epi32(_s0,_sm));
642                     }
643                 }
644                 #elif CV_NEON
645                 float32x4_t v_scale = vdupq_n_f32((float)_scale);
646                 for( ; i <= width-8; i+=8 )
647                 {
648                     int32x4_t v_s0 = vaddq_s32(vld1q_s32(SUM + i), vld1q_s32(Sp + i));
649                     int32x4_t v_s01 = vaddq_s32(vld1q_s32(SUM + i + 4), vld1q_s32(Sp + i + 4));
650 
651                     uint32x4_t v_s0d = cv_vrndq_u32_f32(vmulq_f32(vcvtq_f32_s32(v_s0), v_scale));
652                     uint32x4_t v_s01d = cv_vrndq_u32_f32(vmulq_f32(vcvtq_f32_s32(v_s01), v_scale));
653                     vst1q_u16(D + i, vcombine_u16(vqmovn_u32(v_s0d), vqmovn_u32(v_s01d)));
654 
655                     vst1q_s32(SUM + i, vsubq_s32(v_s0, vld1q_s32(Sm + i)));
656                     vst1q_s32(SUM + i + 4, vsubq_s32(v_s01, vld1q_s32(Sm + i + 4)));
657                 }
658                 #endif
659                 for( ; i < width; i++ )
660                 {
661                     int s0 = SUM[i] + Sp[i];
662                     D[i] = saturate_cast<ushort>(s0*_scale);
663                     SUM[i] = s0 - Sm[i];
664                 }
665             }
666             else
667             {
668                 i = 0;
669                 #if  CV_SSE2
670                 if(haveSSE2)
671                 {
672                     const __m128i delta0 = _mm_set1_epi32(0x8000);
673                     const __m128i delta1 = _mm_set1_epi32(0x80008000);
674 
675                     for( ; i < width-4; i+=4 )
676                     {
677                         __m128i _sm   = _mm_loadu_si128((const __m128i*)(Sm+i));
678                         __m128i _s0   = _mm_add_epi32(_mm_loadu_si128((const __m128i*)(SUM+i)),
679                                                       _mm_loadu_si128((const __m128i*)(Sp+i)));
680 
681                         __m128i _res = _mm_sub_epi32(_s0, delta0);
682                         _res = _mm_add_epi16(_mm_packs_epi32(_res, _res), delta1);
683 
684                         _mm_storel_epi64((__m128i*)(D+i), _res);
685                         _mm_storeu_si128((__m128i*)(SUM+i), _mm_sub_epi32(_s0,_sm));
686                     }
687                 }
688                 #elif CV_NEON
689                 for( ; i <= width-8; i+=8 )
690                 {
691                     int32x4_t v_s0 = vaddq_s32(vld1q_s32(SUM + i), vld1q_s32(Sp + i));
692                     int32x4_t v_s01 = vaddq_s32(vld1q_s32(SUM + i + 4), vld1q_s32(Sp + i + 4));
693 
694                     vst1q_u16(D + i, vcombine_u16(vqmovun_s32(v_s0), vqmovun_s32(v_s01)));
695 
696                     vst1q_s32(SUM + i, vsubq_s32(v_s0, vld1q_s32(Sm + i)));
697                     vst1q_s32(SUM + i + 4, vsubq_s32(v_s01, vld1q_s32(Sm + i + 4)));
698                 }
699                 #endif
700 
701                 for( ; i < width; i++ )
702                 {
703                     int s0 = SUM[i] + Sp[i];
704                     D[i] = saturate_cast<ushort>(s0);
705                     SUM[i] = s0 - Sm[i];
706                 }
707             }
708             dst += dststep;
709         }
710     }
711 
712     double scale;
713     int sumCount;
714     std::vector<int> sum;
715 };
716 
717 template<>
718 struct ColumnSum<int, int> :
719         public BaseColumnFilter
720 {
ColumnSumcv::ColumnSum721     ColumnSum( int _ksize, int _anchor, double _scale ) :
722         BaseColumnFilter()
723     {
724         ksize = _ksize;
725         anchor = _anchor;
726         scale = _scale;
727         sumCount = 0;
728     }
729 
resetcv::ColumnSum730     virtual void reset() { sumCount = 0; }
731 
operator ()cv::ColumnSum732     virtual void operator()(const uchar** src, uchar* dst, int dststep, int count, int width)
733     {
734         int i;
735         int* SUM;
736         bool haveScale = scale != 1;
737         double _scale = scale;
738 
739         #if CV_SSE2
740             bool haveSSE2 = checkHardwareSupport(CV_CPU_SSE2);
741         #endif
742 
743         if( width != (int)sum.size() )
744         {
745             sum.resize(width);
746             sumCount = 0;
747         }
748         SUM = &sum[0];
749         if( sumCount == 0 )
750         {
751             memset((void*)SUM, 0, width*sizeof(int));
752             for( ; sumCount < ksize - 1; sumCount++, src++ )
753             {
754                 const int* Sp = (const int*)src[0];
755                 i = 0;
756                 #if CV_SSE2
757                 if(haveSSE2)
758                 {
759                     for( ; i <= width-4; i+=4 )
760                     {
761                         __m128i _sum = _mm_loadu_si128((const __m128i*)(SUM+i));
762                         __m128i _sp = _mm_loadu_si128((const __m128i*)(Sp+i));
763                         _mm_storeu_si128((__m128i*)(SUM+i),_mm_add_epi32(_sum, _sp));
764                     }
765                 }
766                 #elif CV_NEON
767                 for( ; i <= width - 4; i+=4 )
768                     vst1q_s32(SUM + i, vaddq_s32(vld1q_s32(SUM + i), vld1q_s32(Sp + i)));
769                 #endif
770                 for( ; i < width; i++ )
771                     SUM[i] += Sp[i];
772             }
773         }
774         else
775         {
776             CV_Assert( sumCount == ksize-1 );
777             src += ksize-1;
778         }
779 
780         for( ; count--; src++ )
781         {
782             const int* Sp = (const int*)src[0];
783             const int* Sm = (const int*)src[1-ksize];
784             int* D = (int*)dst;
785             if( haveScale )
786             {
787                 i = 0;
788                 #if CV_SSE2
789                 if(haveSSE2)
790                 {
791                     const __m128 scale4 = _mm_set1_ps((float)_scale);
792                     for( ; i <= width-4; i+=4 )
793                     {
794                         __m128i _sm   = _mm_loadu_si128((const __m128i*)(Sm+i));
795 
796                         __m128i _s0  = _mm_add_epi32(_mm_loadu_si128((const __m128i*)(SUM+i)),
797                                                      _mm_loadu_si128((const __m128i*)(Sp+i)));
798 
799                         __m128i _s0T  = _mm_cvtps_epi32(_mm_mul_ps(scale4, _mm_cvtepi32_ps(_s0)));
800 
801                         _mm_storeu_si128((__m128i*)(D+i), _s0T);
802                         _mm_storeu_si128((__m128i*)(SUM+i),_mm_sub_epi32(_s0,_sm));
803                     }
804                 }
805                 #elif CV_NEON
806                 float32x4_t v_scale = vdupq_n_f32((float)_scale);
807                 for( ; i <= width-4; i+=4 )
808                 {
809                     int32x4_t v_s0 = vaddq_s32(vld1q_s32(SUM + i), vld1q_s32(Sp + i));
810 
811                     int32x4_t v_s0d = cv_vrndq_s32_f32(vmulq_f32(vcvtq_f32_s32(v_s0), v_scale));
812                     vst1q_s32(D + i, v_s0d);
813 
814                     vst1q_s32(SUM + i, vsubq_s32(v_s0, vld1q_s32(Sm + i)));
815                 }
816                 #endif
817                 for( ; i < width; i++ )
818                 {
819                     int s0 = SUM[i] + Sp[i];
820                     D[i] = saturate_cast<int>(s0*_scale);
821                     SUM[i] = s0 - Sm[i];
822                 }
823             }
824             else
825             {
826                 i = 0;
827                 #if CV_SSE2
828                 if(haveSSE2)
829                 {
830                     for( ; i <= width-4; i+=4 )
831                     {
832                         __m128i _sm  = _mm_loadu_si128((const __m128i*)(Sm+i));
833                         __m128i _s0  = _mm_add_epi32(_mm_loadu_si128((const __m128i*)(SUM+i)),
834                                                      _mm_loadu_si128((const __m128i*)(Sp+i)));
835 
836                         _mm_storeu_si128((__m128i*)(D+i), _s0);
837                         _mm_storeu_si128((__m128i*)(SUM+i), _mm_sub_epi32(_s0,_sm));
838                     }
839                 }
840                 #elif CV_NEON
841                 for( ; i <= width-4; i+=4 )
842                 {
843                     int32x4_t v_s0 = vaddq_s32(vld1q_s32(SUM + i), vld1q_s32(Sp + i));
844 
845                     vst1q_s32(D + i, v_s0);
846                     vst1q_s32(SUM + i, vsubq_s32(v_s0, vld1q_s32(Sm + i)));
847                 }
848                 #endif
849 
850                 for( ; i < width; i++ )
851                 {
852                     int s0 = SUM[i] + Sp[i];
853                     D[i] = s0;
854                     SUM[i] = s0 - Sm[i];
855                 }
856             }
857             dst += dststep;
858         }
859     }
860 
861     double scale;
862     int sumCount;
863     std::vector<int> sum;
864 };
865 
866 
867 template<>
868 struct ColumnSum<int, float> :
869         public BaseColumnFilter
870 {
ColumnSumcv::ColumnSum871     ColumnSum( int _ksize, int _anchor, double _scale ) :
872         BaseColumnFilter()
873     {
874         ksize = _ksize;
875         anchor = _anchor;
876         scale = _scale;
877         sumCount = 0;
878     }
879 
resetcv::ColumnSum880     virtual void reset() { sumCount = 0; }
881 
operator ()cv::ColumnSum882     virtual void operator()(const uchar** src, uchar* dst, int dststep, int count, int width)
883     {
884         int i;
885         int* SUM;
886         bool haveScale = scale != 1;
887         double _scale = scale;
888 
889         #if CV_SSE2
890         bool haveSSE2 =  checkHardwareSupport(CV_CPU_SSE2);
891         #endif
892 
893         if( width != (int)sum.size() )
894         {
895             sum.resize(width);
896             sumCount = 0;
897         }
898 
899         SUM = &sum[0];
900         if( sumCount == 0 )
901         {
902             memset((void *)SUM, 0, sizeof(int) * width);
903 
904             for( ; sumCount < ksize - 1; sumCount++, src++ )
905             {
906                 const int* Sp = (const int*)src[0];
907                 i = 0;
908 
909                 #if CV_SSE2
910                 if(haveSSE2)
911                 {
912                     for( ; i < width-4; i+=4 )
913                     {
914                         __m128i _sum = _mm_loadu_si128((const __m128i*)(SUM+i));
915                         __m128i _sp = _mm_loadu_si128((const __m128i*)(Sp+i));
916                         _mm_storeu_si128((__m128i*)(SUM+i), _mm_add_epi32(_sum, _sp));
917                     }
918                 }
919                 #elif CV_NEON
920                 for( ; i <= width - 4; i+=4 )
921                     vst1q_s32(SUM + i, vaddq_s32(vld1q_s32(SUM + i), vld1q_s32(Sp + i)));
922                 #endif
923 
924                 for( ; i < width; i++ )
925                     SUM[i] += Sp[i];
926             }
927         }
928         else
929         {
930             CV_Assert( sumCount == ksize-1 );
931             src += ksize-1;
932         }
933 
934         for( ; count--; src++ )
935         {
936             const int * Sp = (const int*)src[0];
937             const int * Sm = (const int*)src[1-ksize];
938             float* D = (float*)dst;
939             if( haveScale )
940             {
941                 i = 0;
942 
943                 #if CV_SSE2
944                 if(haveSSE2)
945                 {
946                     const __m128 scale4 = _mm_set1_ps((float)_scale);
947 
948                     for( ; i < width-4; i+=4)
949                     {
950                         __m128i _sm   = _mm_loadu_si128((const __m128i*)(Sm+i));
951                         __m128i _s0   = _mm_add_epi32(_mm_loadu_si128((const __m128i*)(SUM+i)),
952                                                       _mm_loadu_si128((const __m128i*)(Sp+i)));
953 
954                         _mm_storeu_ps(D+i, _mm_mul_ps(scale4, _mm_cvtepi32_ps(_s0)));
955                         _mm_storeu_si128((__m128i*)(SUM+i), _mm_sub_epi32(_s0,_sm));
956                     }
957                 }
958                 #elif CV_NEON
959                 float32x4_t v_scale = vdupq_n_f32((float)_scale);
960                 for( ; i <= width-8; i+=8 )
961                 {
962                     int32x4_t v_s0 = vaddq_s32(vld1q_s32(SUM + i), vld1q_s32(Sp + i));
963                     int32x4_t v_s01 = vaddq_s32(vld1q_s32(SUM + i + 4), vld1q_s32(Sp + i + 4));
964 
965                     vst1q_f32(D + i, vmulq_f32(vcvtq_f32_s32(v_s0), v_scale));
966                     vst1q_f32(D + i + 4, vmulq_f32(vcvtq_f32_s32(v_s01), v_scale));
967 
968                     vst1q_s32(SUM + i, vsubq_s32(v_s0, vld1q_s32(Sm + i)));
969                     vst1q_s32(SUM + i + 4, vsubq_s32(v_s01, vld1q_s32(Sm + i + 4)));
970                 }
971                 #endif
972 
973                 for( ; i < width; i++ )
974                 {
975                     int s0 = SUM[i] + Sp[i];
976                     D[i] = (float)(s0*_scale);
977                     SUM[i] = s0 - Sm[i];
978                 }
979             }
980             else
981             {
982                 i = 0;
983 
984                 #if CV_SSE2
985                 if(haveSSE2)
986                 {
987                     for( ; i < width-4; i+=4)
988                     {
989                         __m128i _sm   = _mm_loadu_si128((const __m128i*)(Sm+i));
990                         __m128i _s0   = _mm_add_epi32(_mm_loadu_si128((const __m128i*)(SUM+i)),
991                                                       _mm_loadu_si128((const __m128i*)(Sp+i)));
992 
993                         _mm_storeu_ps(D+i, _mm_cvtepi32_ps(_s0));
994                         _mm_storeu_si128((__m128i*)(SUM+i), _mm_sub_epi32(_s0,_sm));
995                     }
996                 }
997                 #elif CV_NEON
998                 for( ; i <= width-8; i+=8 )
999                 {
1000                     int32x4_t v_s0 = vaddq_s32(vld1q_s32(SUM + i), vld1q_s32(Sp + i));
1001                     int32x4_t v_s01 = vaddq_s32(vld1q_s32(SUM + i + 4), vld1q_s32(Sp + i + 4));
1002 
1003                     vst1q_f32(D + i, vcvtq_f32_s32(v_s0));
1004                     vst1q_f32(D + i + 4, vcvtq_f32_s32(v_s01));
1005 
1006                     vst1q_s32(SUM + i, vsubq_s32(v_s0, vld1q_s32(Sm + i)));
1007                     vst1q_s32(SUM + i + 4, vsubq_s32(v_s01, vld1q_s32(Sm + i + 4)));
1008                 }
1009                 #endif
1010 
1011                 for( ; i < width; i++ )
1012                 {
1013                     int s0 = SUM[i] + Sp[i];
1014                     D[i] = (float)(s0);
1015                     SUM[i] = s0 - Sm[i];
1016                 }
1017             }
1018             dst += dststep;
1019         }
1020     }
1021 
1022     double scale;
1023     int sumCount;
1024     std::vector<int> sum;
1025 };
1026 
1027 #ifdef HAVE_OPENCL
1028 
1029 #define DIVUP(total, grain) ((total + grain - 1) / (grain))
1030 #define ROUNDUP(sz, n)      ((sz) + (n) - 1 - (((sz) + (n) - 1) % (n)))
1031 
ocl_boxFilter(InputArray _src,OutputArray _dst,int ddepth,Size ksize,Point anchor,int borderType,bool normalize,bool sqr=false)1032 static bool ocl_boxFilter( InputArray _src, OutputArray _dst, int ddepth,
1033                            Size ksize, Point anchor, int borderType, bool normalize, bool sqr = false )
1034 {
1035     const ocl::Device & dev = ocl::Device::getDefault();
1036     int type = _src.type(), sdepth = CV_MAT_DEPTH(type), cn = CV_MAT_CN(type), esz = CV_ELEM_SIZE(type);
1037     bool doubleSupport = dev.doubleFPConfig() > 0;
1038 
1039     if (ddepth < 0)
1040         ddepth = sdepth;
1041 
1042     if (cn > 4 || (!doubleSupport && (sdepth == CV_64F || ddepth == CV_64F)) ||
1043         _src.offset() % esz != 0 || _src.step() % esz != 0)
1044         return false;
1045 
1046     if (anchor.x < 0)
1047         anchor.x = ksize.width / 2;
1048     if (anchor.y < 0)
1049         anchor.y = ksize.height / 2;
1050 
1051     int computeUnits = ocl::Device::getDefault().maxComputeUnits();
1052     float alpha = 1.0f / (ksize.height * ksize.width);
1053     Size size = _src.size(), wholeSize;
1054     bool isolated = (borderType & BORDER_ISOLATED) != 0;
1055     borderType &= ~BORDER_ISOLATED;
1056     int wdepth = std::max(CV_32F, std::max(ddepth, sdepth)),
1057         wtype = CV_MAKE_TYPE(wdepth, cn), dtype = CV_MAKE_TYPE(ddepth, cn);
1058 
1059     const char * const borderMap[] = { "BORDER_CONSTANT", "BORDER_REPLICATE", "BORDER_REFLECT", 0, "BORDER_REFLECT_101" };
1060     size_t globalsize[2] = { size.width, size.height };
1061     size_t localsize_general[2] = { 0, 1 }, * localsize = NULL;
1062 
1063     UMat src = _src.getUMat();
1064     if (!isolated)
1065     {
1066         Point ofs;
1067         src.locateROI(wholeSize, ofs);
1068     }
1069 
1070     int h = isolated ? size.height : wholeSize.height;
1071     int w = isolated ? size.width : wholeSize.width;
1072 
1073     size_t maxWorkItemSizes[32];
1074     ocl::Device::getDefault().maxWorkItemSizes(maxWorkItemSizes);
1075     int tryWorkItems = (int)maxWorkItemSizes[0];
1076 
1077     ocl::Kernel kernel;
1078 
1079     if (dev.isIntel() && !(dev.type() & ocl::Device::TYPE_CPU) &&
1080         ((ksize.width < 5 && ksize.height < 5 && esz <= 4) ||
1081          (ksize.width == 5 && ksize.height == 5 && cn == 1)))
1082     {
1083         if (w < ksize.width || h < ksize.height)
1084             return false;
1085 
1086         // Figure out what vector size to use for loading the pixels.
1087         int pxLoadNumPixels = cn != 1 || size.width % 4 ? 1 : 4;
1088         int pxLoadVecSize = cn * pxLoadNumPixels;
1089 
1090         // Figure out how many pixels per work item to compute in X and Y
1091         // directions.  Too many and we run out of registers.
1092         int pxPerWorkItemX = 1, pxPerWorkItemY = 1;
1093         if (cn <= 2 && ksize.width <= 4 && ksize.height <= 4)
1094         {
1095             pxPerWorkItemX = size.width % 8 ? size.width % 4 ? size.width % 2 ? 1 : 2 : 4 : 8;
1096             pxPerWorkItemY = size.height % 2 ? 1 : 2;
1097         }
1098         else if (cn < 4 || (ksize.width <= 4 && ksize.height <= 4))
1099         {
1100             pxPerWorkItemX = size.width % 2 ? 1 : 2;
1101             pxPerWorkItemY = size.height % 2 ? 1 : 2;
1102         }
1103         globalsize[0] = size.width / pxPerWorkItemX;
1104         globalsize[1] = size.height / pxPerWorkItemY;
1105 
1106         // Need some padding in the private array for pixels
1107         int privDataWidth = ROUNDUP(pxPerWorkItemX + ksize.width - 1, pxLoadNumPixels);
1108 
1109         // Make the global size a nice round number so the runtime can pick
1110         // from reasonable choices for the workgroup size
1111         const int wgRound = 256;
1112         globalsize[0] = ROUNDUP(globalsize[0], wgRound);
1113 
1114         char build_options[1024], cvt[2][40];
1115         sprintf(build_options, "-D cn=%d "
1116                 "-D ANCHOR_X=%d -D ANCHOR_Y=%d -D KERNEL_SIZE_X=%d -D KERNEL_SIZE_Y=%d "
1117                 "-D PX_LOAD_VEC_SIZE=%d -D PX_LOAD_NUM_PX=%d "
1118                 "-D PX_PER_WI_X=%d -D PX_PER_WI_Y=%d -D PRIV_DATA_WIDTH=%d -D %s -D %s "
1119                 "-D PX_LOAD_X_ITERATIONS=%d -D PX_LOAD_Y_ITERATIONS=%d "
1120                 "-D srcT=%s -D srcT1=%s -D dstT=%s -D dstT1=%s -D WT=%s -D WT1=%s "
1121                 "-D convertToWT=%s -D convertToDstT=%s%s%s -D PX_LOAD_FLOAT_VEC_CONV=convert_%s -D OP_BOX_FILTER",
1122                 cn, anchor.x, anchor.y, ksize.width, ksize.height,
1123                 pxLoadVecSize, pxLoadNumPixels,
1124                 pxPerWorkItemX, pxPerWorkItemY, privDataWidth, borderMap[borderType],
1125                 isolated ? "BORDER_ISOLATED" : "NO_BORDER_ISOLATED",
1126                 privDataWidth / pxLoadNumPixels, pxPerWorkItemY + ksize.height - 1,
1127                 ocl::typeToStr(type), ocl::typeToStr(sdepth), ocl::typeToStr(dtype),
1128                 ocl::typeToStr(ddepth), ocl::typeToStr(wtype), ocl::typeToStr(wdepth),
1129                 ocl::convertTypeStr(sdepth, wdepth, cn, cvt[0]),
1130                 ocl::convertTypeStr(wdepth, ddepth, cn, cvt[1]),
1131                 normalize ? " -D NORMALIZE" : "", sqr ? " -D SQR" : "",
1132                 ocl::typeToStr(CV_MAKE_TYPE(wdepth, pxLoadVecSize)) //PX_LOAD_FLOAT_VEC_CONV
1133                 );
1134 
1135 
1136         if (!kernel.create("filterSmall", cv::ocl::imgproc::filterSmall_oclsrc, build_options))
1137             return false;
1138     }
1139     else
1140     {
1141         localsize = localsize_general;
1142         for ( ; ; )
1143         {
1144             int BLOCK_SIZE_X = tryWorkItems, BLOCK_SIZE_Y = std::min(ksize.height * 10, size.height);
1145 
1146             while (BLOCK_SIZE_X > 32 && BLOCK_SIZE_X >= ksize.width * 2 && BLOCK_SIZE_X > size.width * 2)
1147                 BLOCK_SIZE_X /= 2;
1148             while (BLOCK_SIZE_Y < BLOCK_SIZE_X / 8 && BLOCK_SIZE_Y * computeUnits * 32 < size.height)
1149                 BLOCK_SIZE_Y *= 2;
1150 
1151             if (ksize.width > BLOCK_SIZE_X || w < ksize.width || h < ksize.height)
1152                 return false;
1153 
1154             char cvt[2][50];
1155             String opts = format("-D LOCAL_SIZE_X=%d -D BLOCK_SIZE_Y=%d -D ST=%s -D DT=%s -D WT=%s -D convertToDT=%s -D convertToWT=%s"
1156                                  " -D ANCHOR_X=%d -D ANCHOR_Y=%d -D KERNEL_SIZE_X=%d -D KERNEL_SIZE_Y=%d -D %s%s%s%s%s"
1157                                  " -D ST1=%s -D DT1=%s -D cn=%d",
1158                                  BLOCK_SIZE_X, BLOCK_SIZE_Y, ocl::typeToStr(type), ocl::typeToStr(CV_MAKE_TYPE(ddepth, cn)),
1159                                  ocl::typeToStr(CV_MAKE_TYPE(wdepth, cn)),
1160                                  ocl::convertTypeStr(wdepth, ddepth, cn, cvt[0]),
1161                                  ocl::convertTypeStr(sdepth, wdepth, cn, cvt[1]),
1162                                  anchor.x, anchor.y, ksize.width, ksize.height, borderMap[borderType],
1163                                  isolated ? " -D BORDER_ISOLATED" : "", doubleSupport ? " -D DOUBLE_SUPPORT" : "",
1164                                  normalize ? " -D NORMALIZE" : "", sqr ? " -D SQR" : "",
1165                                  ocl::typeToStr(sdepth), ocl::typeToStr(ddepth), cn);
1166 
1167             localsize[0] = BLOCK_SIZE_X;
1168             globalsize[0] = DIVUP(size.width, BLOCK_SIZE_X - (ksize.width - 1)) * BLOCK_SIZE_X;
1169             globalsize[1] = DIVUP(size.height, BLOCK_SIZE_Y);
1170 
1171             kernel.create("boxFilter", cv::ocl::imgproc::boxFilter_oclsrc, opts);
1172             if (kernel.empty())
1173                 return false;
1174 
1175             size_t kernelWorkGroupSize = kernel.workGroupSize();
1176             if (localsize[0] <= kernelWorkGroupSize)
1177                 break;
1178             if (BLOCK_SIZE_X < (int)kernelWorkGroupSize)
1179                 return false;
1180 
1181             tryWorkItems = (int)kernelWorkGroupSize;
1182         }
1183     }
1184 
1185     _dst.create(size, CV_MAKETYPE(ddepth, cn));
1186     UMat dst = _dst.getUMat();
1187 
1188     int idxArg = kernel.set(0, ocl::KernelArg::PtrReadOnly(src));
1189     idxArg = kernel.set(idxArg, (int)src.step);
1190     int srcOffsetX = (int)((src.offset % src.step) / src.elemSize());
1191     int srcOffsetY = (int)(src.offset / src.step);
1192     int srcEndX = isolated ? srcOffsetX + size.width : wholeSize.width;
1193     int srcEndY = isolated ? srcOffsetY + size.height : wholeSize.height;
1194     idxArg = kernel.set(idxArg, srcOffsetX);
1195     idxArg = kernel.set(idxArg, srcOffsetY);
1196     idxArg = kernel.set(idxArg, srcEndX);
1197     idxArg = kernel.set(idxArg, srcEndY);
1198     idxArg = kernel.set(idxArg, ocl::KernelArg::WriteOnly(dst));
1199     if (normalize)
1200         idxArg = kernel.set(idxArg, (float)alpha);
1201 
1202     return kernel.run(2, globalsize, localsize, false);
1203 }
1204 
1205 #undef ROUNDUP
1206 
1207 #endif
1208 
1209 }
1210 
1211 
getRowSumFilter(int srcType,int sumType,int ksize,int anchor)1212 cv::Ptr<cv::BaseRowFilter> cv::getRowSumFilter(int srcType, int sumType, int ksize, int anchor)
1213 {
1214     int sdepth = CV_MAT_DEPTH(srcType), ddepth = CV_MAT_DEPTH(sumType);
1215     CV_Assert( CV_MAT_CN(sumType) == CV_MAT_CN(srcType) );
1216 
1217     if( anchor < 0 )
1218         anchor = ksize/2;
1219 
1220     if( sdepth == CV_8U && ddepth == CV_32S )
1221         return makePtr<RowSum<uchar, int> >(ksize, anchor);
1222     if( sdepth == CV_8U && ddepth == CV_64F )
1223         return makePtr<RowSum<uchar, double> >(ksize, anchor);
1224     if( sdepth == CV_16U && ddepth == CV_32S )
1225         return makePtr<RowSum<ushort, int> >(ksize, anchor);
1226     if( sdepth == CV_16U && ddepth == CV_64F )
1227         return makePtr<RowSum<ushort, double> >(ksize, anchor);
1228     if( sdepth == CV_16S && ddepth == CV_32S )
1229         return makePtr<RowSum<short, int> >(ksize, anchor);
1230     if( sdepth == CV_32S && ddepth == CV_32S )
1231         return makePtr<RowSum<int, int> >(ksize, anchor);
1232     if( sdepth == CV_16S && ddepth == CV_64F )
1233         return makePtr<RowSum<short, double> >(ksize, anchor);
1234     if( sdepth == CV_32F && ddepth == CV_64F )
1235         return makePtr<RowSum<float, double> >(ksize, anchor);
1236     if( sdepth == CV_64F && ddepth == CV_64F )
1237         return makePtr<RowSum<double, double> >(ksize, anchor);
1238 
1239     CV_Error_( CV_StsNotImplemented,
1240         ("Unsupported combination of source format (=%d), and buffer format (=%d)",
1241         srcType, sumType));
1242 
1243     return Ptr<BaseRowFilter>();
1244 }
1245 
1246 
getColumnSumFilter(int sumType,int dstType,int ksize,int anchor,double scale)1247 cv::Ptr<cv::BaseColumnFilter> cv::getColumnSumFilter(int sumType, int dstType, int ksize,
1248                                                      int anchor, double scale)
1249 {
1250     int sdepth = CV_MAT_DEPTH(sumType), ddepth = CV_MAT_DEPTH(dstType);
1251     CV_Assert( CV_MAT_CN(sumType) == CV_MAT_CN(dstType) );
1252 
1253     if( anchor < 0 )
1254         anchor = ksize/2;
1255 
1256     if( ddepth == CV_8U && sdepth == CV_32S )
1257         return makePtr<ColumnSum<int, uchar> >(ksize, anchor, scale);
1258     if( ddepth == CV_8U && sdepth == CV_64F )
1259         return makePtr<ColumnSum<double, uchar> >(ksize, anchor, scale);
1260     if( ddepth == CV_16U && sdepth == CV_32S )
1261         return makePtr<ColumnSum<int, ushort> >(ksize, anchor, scale);
1262     if( ddepth == CV_16U && sdepth == CV_64F )
1263         return makePtr<ColumnSum<double, ushort> >(ksize, anchor, scale);
1264     if( ddepth == CV_16S && sdepth == CV_32S )
1265         return makePtr<ColumnSum<int, short> >(ksize, anchor, scale);
1266     if( ddepth == CV_16S && sdepth == CV_64F )
1267         return makePtr<ColumnSum<double, short> >(ksize, anchor, scale);
1268     if( ddepth == CV_32S && sdepth == CV_32S )
1269         return makePtr<ColumnSum<int, int> >(ksize, anchor, scale);
1270     if( ddepth == CV_32F && sdepth == CV_32S )
1271         return makePtr<ColumnSum<int, float> >(ksize, anchor, scale);
1272     if( ddepth == CV_32F && sdepth == CV_64F )
1273         return makePtr<ColumnSum<double, float> >(ksize, anchor, scale);
1274     if( ddepth == CV_64F && sdepth == CV_32S )
1275         return makePtr<ColumnSum<int, double> >(ksize, anchor, scale);
1276     if( ddepth == CV_64F && sdepth == CV_64F )
1277         return makePtr<ColumnSum<double, double> >(ksize, anchor, scale);
1278 
1279     CV_Error_( CV_StsNotImplemented,
1280         ("Unsupported combination of sum format (=%d), and destination format (=%d)",
1281         sumType, dstType));
1282 
1283     return Ptr<BaseColumnFilter>();
1284 }
1285 
1286 
createBoxFilter(int srcType,int dstType,Size ksize,Point anchor,bool normalize,int borderType)1287 cv::Ptr<cv::FilterEngine> cv::createBoxFilter( int srcType, int dstType, Size ksize,
1288                     Point anchor, bool normalize, int borderType )
1289 {
1290     int sdepth = CV_MAT_DEPTH(srcType);
1291     int cn = CV_MAT_CN(srcType), sumType = CV_64F;
1292     if( sdepth <= CV_32S && (!normalize ||
1293         ksize.width*ksize.height <= (sdepth == CV_8U ? (1<<23) :
1294             sdepth == CV_16U ? (1 << 15) : (1 << 16))) )
1295         sumType = CV_32S;
1296     sumType = CV_MAKETYPE( sumType, cn );
1297 
1298     Ptr<BaseRowFilter> rowFilter = getRowSumFilter(srcType, sumType, ksize.width, anchor.x );
1299     Ptr<BaseColumnFilter> columnFilter = getColumnSumFilter(sumType,
1300         dstType, ksize.height, anchor.y, normalize ? 1./(ksize.width*ksize.height) : 1);
1301 
1302     return makePtr<FilterEngine>(Ptr<BaseFilter>(), rowFilter, columnFilter,
1303            srcType, dstType, sumType, borderType );
1304 }
1305 
1306 
boxFilter(InputArray _src,OutputArray _dst,int ddepth,Size ksize,Point anchor,bool normalize,int borderType)1307 void cv::boxFilter( InputArray _src, OutputArray _dst, int ddepth,
1308                 Size ksize, Point anchor,
1309                 bool normalize, int borderType )
1310 {
1311     CV_OCL_RUN(_dst.isUMat(), ocl_boxFilter(_src, _dst, ddepth, ksize, anchor, borderType, normalize))
1312 
1313     Mat src = _src.getMat();
1314     int stype = src.type(), sdepth = CV_MAT_DEPTH(stype), cn = CV_MAT_CN(stype);
1315     if( ddepth < 0 )
1316         ddepth = sdepth;
1317     _dst.create( src.size(), CV_MAKETYPE(ddepth, cn) );
1318     Mat dst = _dst.getMat();
1319     if( borderType != BORDER_CONSTANT && normalize && (borderType & BORDER_ISOLATED) != 0 )
1320     {
1321         if( src.rows == 1 )
1322             ksize.height = 1;
1323         if( src.cols == 1 )
1324             ksize.width = 1;
1325     }
1326 #ifdef HAVE_TEGRA_OPTIMIZATION
1327     if ( tegra::useTegra() && tegra::box(src, dst, ksize, anchor, normalize, borderType) )
1328         return;
1329 #endif
1330 
1331 #if defined(HAVE_IPP)
1332     CV_IPP_CHECK()
1333     {
1334         int ippBorderType = borderType & ~BORDER_ISOLATED;
1335         Point ocvAnchor, ippAnchor;
1336         ocvAnchor.x = anchor.x < 0 ? ksize.width / 2 : anchor.x;
1337         ocvAnchor.y = anchor.y < 0 ? ksize.height / 2 : anchor.y;
1338         ippAnchor.x = ksize.width / 2 - (ksize.width % 2 == 0 ? 1 : 0);
1339         ippAnchor.y = ksize.height / 2 - (ksize.height % 2 == 0 ? 1 : 0);
1340 
1341         if (normalize && !src.isSubmatrix() && ddepth == sdepth &&
1342             (/*ippBorderType == BORDER_REPLICATE ||*/ /* returns ippStsStepErr: Step value is not valid */
1343              ippBorderType == BORDER_CONSTANT) && ocvAnchor == ippAnchor &&
1344              dst.cols != ksize.width && dst.rows != ksize.height) // returns ippStsMaskSizeErr: mask has an illegal value
1345         {
1346             Ipp32s bufSize = 0;
1347             IppiSize roiSize = { dst.cols, dst.rows }, maskSize = { ksize.width, ksize.height };
1348 
1349 #define IPP_FILTER_BOX_BORDER(ippType, ippDataType, flavor) \
1350             do \
1351             { \
1352                 if (ippiFilterBoxBorderGetBufferSize(roiSize, maskSize, ippDataType, cn, &bufSize) >= 0) \
1353                 { \
1354                     Ipp8u * buffer = ippsMalloc_8u(bufSize); \
1355                     ippType borderValue[4] = { 0, 0, 0, 0 }; \
1356                     ippBorderType = ippBorderType == BORDER_CONSTANT ? ippBorderConst : ippBorderRepl; \
1357                     IppStatus status = ippiFilterBoxBorder_##flavor(src.ptr<ippType>(), (int)src.step, dst.ptr<ippType>(), \
1358                                                                     (int)dst.step, roiSize, maskSize, \
1359                                                                     (IppiBorderType)ippBorderType, borderValue, buffer); \
1360                     ippsFree(buffer); \
1361                     if (status >= 0) \
1362                     { \
1363                         CV_IMPL_ADD(CV_IMPL_IPP); \
1364                         return; \
1365                     } \
1366                 } \
1367                 setIppErrorStatus(); \
1368             } while ((void)0, 0)
1369 
1370             if (stype == CV_8UC1)
1371                 IPP_FILTER_BOX_BORDER(Ipp8u, ipp8u, 8u_C1R);
1372             else if (stype == CV_8UC3)
1373                 IPP_FILTER_BOX_BORDER(Ipp8u, ipp8u, 8u_C3R);
1374             else if (stype == CV_8UC4)
1375                 IPP_FILTER_BOX_BORDER(Ipp8u, ipp8u, 8u_C4R);
1376 
1377             // Oct 2014: performance with BORDER_CONSTANT
1378             //else if (stype == CV_16UC1)
1379             //    IPP_FILTER_BOX_BORDER(Ipp16u, ipp16u, 16u_C1R);
1380             else if (stype == CV_16UC3)
1381                 IPP_FILTER_BOX_BORDER(Ipp16u, ipp16u, 16u_C3R);
1382             else if (stype == CV_16UC4)
1383                 IPP_FILTER_BOX_BORDER(Ipp16u, ipp16u, 16u_C4R);
1384 
1385             // Oct 2014: performance with BORDER_CONSTANT
1386             //else if (stype == CV_16SC1)
1387             //    IPP_FILTER_BOX_BORDER(Ipp16s, ipp16s, 16s_C1R);
1388             else if (stype == CV_16SC3)
1389                 IPP_FILTER_BOX_BORDER(Ipp16s, ipp16s, 16s_C3R);
1390             else if (stype == CV_16SC4)
1391                 IPP_FILTER_BOX_BORDER(Ipp16s, ipp16s, 16s_C4R);
1392 
1393             else if (stype == CV_32FC1)
1394                 IPP_FILTER_BOX_BORDER(Ipp32f, ipp32f, 32f_C1R);
1395             else if (stype == CV_32FC3)
1396                 IPP_FILTER_BOX_BORDER(Ipp32f, ipp32f, 32f_C3R);
1397             else if (stype == CV_32FC4)
1398                 IPP_FILTER_BOX_BORDER(Ipp32f, ipp32f, 32f_C4R);
1399         }
1400 #undef IPP_FILTER_BOX_BORDER
1401     }
1402 #endif
1403 
1404     Ptr<FilterEngine> f = createBoxFilter( src.type(), dst.type(),
1405                         ksize, anchor, normalize, borderType );
1406     f->apply( src, dst );
1407 }
1408 
blur(InputArray src,OutputArray dst,Size ksize,Point anchor,int borderType)1409 void cv::blur( InputArray src, OutputArray dst,
1410            Size ksize, Point anchor, int borderType )
1411 {
1412     boxFilter( src, dst, -1, ksize, anchor, true, borderType );
1413 }
1414 
1415 
1416 /****************************************************************************************\
1417                                     Squared Box Filter
1418 \****************************************************************************************/
1419 
1420 namespace cv
1421 {
1422 
1423 template<typename T, typename ST>
1424 struct SqrRowSum :
1425         public BaseRowFilter
1426 {
SqrRowSumcv::SqrRowSum1427     SqrRowSum( int _ksize, int _anchor ) :
1428         BaseRowFilter()
1429     {
1430         ksize = _ksize;
1431         anchor = _anchor;
1432     }
1433 
operator ()cv::SqrRowSum1434     virtual void operator()(const uchar* src, uchar* dst, int width, int cn)
1435     {
1436         const T* S = (const T*)src;
1437         ST* D = (ST*)dst;
1438         int i = 0, k, ksz_cn = ksize*cn;
1439 
1440         width = (width - 1)*cn;
1441         for( k = 0; k < cn; k++, S++, D++ )
1442         {
1443             ST s = 0;
1444             for( i = 0; i < ksz_cn; i += cn )
1445             {
1446                 ST val = (ST)S[i];
1447                 s += val*val;
1448             }
1449             D[0] = s;
1450             for( i = 0; i < width; i += cn )
1451             {
1452                 ST val0 = (ST)S[i], val1 = (ST)S[i + ksz_cn];
1453                 s += val1*val1 - val0*val0;
1454                 D[i+cn] = s;
1455             }
1456         }
1457     }
1458 };
1459 
getSqrRowSumFilter(int srcType,int sumType,int ksize,int anchor)1460 static Ptr<BaseRowFilter> getSqrRowSumFilter(int srcType, int sumType, int ksize, int anchor)
1461 {
1462     int sdepth = CV_MAT_DEPTH(srcType), ddepth = CV_MAT_DEPTH(sumType);
1463     CV_Assert( CV_MAT_CN(sumType) == CV_MAT_CN(srcType) );
1464 
1465     if( anchor < 0 )
1466         anchor = ksize/2;
1467 
1468     if( sdepth == CV_8U && ddepth == CV_32S )
1469         return makePtr<SqrRowSum<uchar, int> >(ksize, anchor);
1470     if( sdepth == CV_8U && ddepth == CV_64F )
1471         return makePtr<SqrRowSum<uchar, double> >(ksize, anchor);
1472     if( sdepth == CV_16U && ddepth == CV_64F )
1473         return makePtr<SqrRowSum<ushort, double> >(ksize, anchor);
1474     if( sdepth == CV_16S && ddepth == CV_64F )
1475         return makePtr<SqrRowSum<short, double> >(ksize, anchor);
1476     if( sdepth == CV_32F && ddepth == CV_64F )
1477         return makePtr<SqrRowSum<float, double> >(ksize, anchor);
1478     if( sdepth == CV_64F && ddepth == CV_64F )
1479         return makePtr<SqrRowSum<double, double> >(ksize, anchor);
1480 
1481     CV_Error_( CV_StsNotImplemented,
1482               ("Unsupported combination of source format (=%d), and buffer format (=%d)",
1483                srcType, sumType));
1484 
1485     return Ptr<BaseRowFilter>();
1486 }
1487 
1488 }
1489 
sqrBoxFilter(InputArray _src,OutputArray _dst,int ddepth,Size ksize,Point anchor,bool normalize,int borderType)1490 void cv::sqrBoxFilter( InputArray _src, OutputArray _dst, int ddepth,
1491                        Size ksize, Point anchor,
1492                        bool normalize, int borderType )
1493 {
1494     int srcType = _src.type(), sdepth = CV_MAT_DEPTH(srcType), cn = CV_MAT_CN(srcType);
1495     Size size = _src.size();
1496 
1497     if( ddepth < 0 )
1498         ddepth = sdepth < CV_32F ? CV_32F : CV_64F;
1499 
1500     if( borderType != BORDER_CONSTANT && normalize )
1501     {
1502         if( size.height == 1 )
1503             ksize.height = 1;
1504         if( size.width == 1 )
1505             ksize.width = 1;
1506     }
1507 
1508     CV_OCL_RUN(_dst.isUMat() && _src.dims() <= 2,
1509                ocl_boxFilter(_src, _dst, ddepth, ksize, anchor, borderType, normalize, true))
1510 
1511     int sumDepth = CV_64F;
1512     if( sdepth == CV_8U )
1513         sumDepth = CV_32S;
1514     int sumType = CV_MAKETYPE( sumDepth, cn ), dstType = CV_MAKETYPE(ddepth, cn);
1515 
1516     Mat src = _src.getMat();
1517     _dst.create( size, dstType );
1518     Mat dst = _dst.getMat();
1519 
1520     Ptr<BaseRowFilter> rowFilter = getSqrRowSumFilter(srcType, sumType, ksize.width, anchor.x );
1521     Ptr<BaseColumnFilter> columnFilter = getColumnSumFilter(sumType,
1522                                                             dstType, ksize.height, anchor.y,
1523                                                             normalize ? 1./(ksize.width*ksize.height) : 1);
1524 
1525     Ptr<FilterEngine> f = makePtr<FilterEngine>(Ptr<BaseFilter>(), rowFilter, columnFilter,
1526                                                 srcType, dstType, sumType, borderType );
1527     f->apply( src, dst );
1528 }
1529 
1530 
1531 /****************************************************************************************\
1532                                      Gaussian Blur
1533 \****************************************************************************************/
1534 
getGaussianKernel(int n,double sigma,int ktype)1535 cv::Mat cv::getGaussianKernel( int n, double sigma, int ktype )
1536 {
1537     const int SMALL_GAUSSIAN_SIZE = 7;
1538     static const float small_gaussian_tab[][SMALL_GAUSSIAN_SIZE] =
1539     {
1540         {1.f},
1541         {0.25f, 0.5f, 0.25f},
1542         {0.0625f, 0.25f, 0.375f, 0.25f, 0.0625f},
1543         {0.03125f, 0.109375f, 0.21875f, 0.28125f, 0.21875f, 0.109375f, 0.03125f}
1544     };
1545 
1546     const float* fixed_kernel = n % 2 == 1 && n <= SMALL_GAUSSIAN_SIZE && sigma <= 0 ?
1547         small_gaussian_tab[n>>1] : 0;
1548 
1549     CV_Assert( ktype == CV_32F || ktype == CV_64F );
1550     Mat kernel(n, 1, ktype);
1551     float* cf = kernel.ptr<float>();
1552     double* cd = kernel.ptr<double>();
1553 
1554     double sigmaX = sigma > 0 ? sigma : ((n-1)*0.5 - 1)*0.3 + 0.8;
1555     double scale2X = -0.5/(sigmaX*sigmaX);
1556     double sum = 0;
1557 
1558     int i;
1559     for( i = 0; i < n; i++ )
1560     {
1561         double x = i - (n-1)*0.5;
1562         double t = fixed_kernel ? (double)fixed_kernel[i] : std::exp(scale2X*x*x);
1563         if( ktype == CV_32F )
1564         {
1565             cf[i] = (float)t;
1566             sum += cf[i];
1567         }
1568         else
1569         {
1570             cd[i] = t;
1571             sum += cd[i];
1572         }
1573     }
1574 
1575     sum = 1./sum;
1576     for( i = 0; i < n; i++ )
1577     {
1578         if( ktype == CV_32F )
1579             cf[i] = (float)(cf[i]*sum);
1580         else
1581             cd[i] *= sum;
1582     }
1583 
1584     return kernel;
1585 }
1586 
1587 namespace cv {
1588 
createGaussianKernels(Mat & kx,Mat & ky,int type,Size ksize,double sigma1,double sigma2)1589 static void createGaussianKernels( Mat & kx, Mat & ky, int type, Size ksize,
1590                                    double sigma1, double sigma2 )
1591 {
1592     int depth = CV_MAT_DEPTH(type);
1593     if( sigma2 <= 0 )
1594         sigma2 = sigma1;
1595 
1596     // automatic detection of kernel size from sigma
1597     if( ksize.width <= 0 && sigma1 > 0 )
1598         ksize.width = cvRound(sigma1*(depth == CV_8U ? 3 : 4)*2 + 1)|1;
1599     if( ksize.height <= 0 && sigma2 > 0 )
1600         ksize.height = cvRound(sigma2*(depth == CV_8U ? 3 : 4)*2 + 1)|1;
1601 
1602     CV_Assert( ksize.width > 0 && ksize.width % 2 == 1 &&
1603         ksize.height > 0 && ksize.height % 2 == 1 );
1604 
1605     sigma1 = std::max( sigma1, 0. );
1606     sigma2 = std::max( sigma2, 0. );
1607 
1608     kx = getGaussianKernel( ksize.width, sigma1, std::max(depth, CV_32F) );
1609     if( ksize.height == ksize.width && std::abs(sigma1 - sigma2) < DBL_EPSILON )
1610         ky = kx;
1611     else
1612         ky = getGaussianKernel( ksize.height, sigma2, std::max(depth, CV_32F) );
1613 }
1614 
1615 }
1616 
createGaussianFilter(int type,Size ksize,double sigma1,double sigma2,int borderType)1617 cv::Ptr<cv::FilterEngine> cv::createGaussianFilter( int type, Size ksize,
1618                                         double sigma1, double sigma2,
1619                                         int borderType )
1620 {
1621     Mat kx, ky;
1622     createGaussianKernels(kx, ky, type, ksize, sigma1, sigma2);
1623 
1624     return createSeparableLinearFilter( type, type, kx, ky, Point(-1,-1), 0, borderType );
1625 }
1626 
1627 
GaussianBlur(InputArray _src,OutputArray _dst,Size ksize,double sigma1,double sigma2,int borderType)1628 void cv::GaussianBlur( InputArray _src, OutputArray _dst, Size ksize,
1629                    double sigma1, double sigma2,
1630                    int borderType )
1631 {
1632     int type = _src.type();
1633     Size size = _src.size();
1634     _dst.create( size, type );
1635 
1636     if( borderType != BORDER_CONSTANT && (borderType & BORDER_ISOLATED) != 0 )
1637     {
1638         if( size.height == 1 )
1639             ksize.height = 1;
1640         if( size.width == 1 )
1641             ksize.width = 1;
1642     }
1643 
1644     if( ksize.width == 1 && ksize.height == 1 )
1645     {
1646         _src.copyTo(_dst);
1647         return;
1648     }
1649 
1650 #ifdef HAVE_TEGRA_OPTIMIZATION
1651     Mat src = _src.getMat();
1652     Mat dst = _dst.getMat();
1653     if(sigma1 == 0 && sigma2 == 0 && tegra::useTegra() && tegra::gaussian(src, dst, ksize, borderType))
1654         return;
1655 #endif
1656 
1657 #if IPP_VERSION_X100 >= 801 && 0 // these functions are slower in IPP 8.1
1658     CV_IPP_CHECK()
1659     {
1660         int depth = CV_MAT_DEPTH(type), cn = CV_MAT_CN(type);
1661 
1662         if ((depth == CV_8U || depth == CV_16U || depth == CV_16S || depth == CV_32F) && (cn == 1 || cn == 3) &&
1663                 sigma1 == sigma2 && ksize.width == ksize.height && sigma1 != 0.0 )
1664         {
1665             IppiBorderType ippBorder = ippiGetBorderType(borderType);
1666             if (ippBorderConst == ippBorder || ippBorderRepl == ippBorder)
1667             {
1668                 Mat src = _src.getMat(), dst = _dst.getMat();
1669                 IppiSize roiSize = { src.cols, src.rows };
1670                 IppDataType dataType = ippiGetDataType(depth);
1671                 Ipp32s specSize = 0, bufferSize = 0;
1672 
1673                 if (ippiFilterGaussianGetBufferSize(roiSize, (Ipp32u)ksize.width, dataType, cn, &specSize, &bufferSize) >= 0)
1674                 {
1675                     IppFilterGaussianSpec * pSpec = (IppFilterGaussianSpec *)ippMalloc(specSize);
1676                     Ipp8u * pBuffer = (Ipp8u*)ippMalloc(bufferSize);
1677 
1678                     if (ippiFilterGaussianInit(roiSize, (Ipp32u)ksize.width, (Ipp32f)sigma1, ippBorder, dataType, 1, pSpec, pBuffer) >= 0)
1679                     {
1680 #define IPP_FILTER_GAUSS(ippfavor, ippcn) \
1681         do \
1682         { \
1683             typedef Ipp##ippfavor ippType; \
1684             ippType borderValues[] = { 0, 0, 0 }; \
1685             IppStatus status = ippcn == 1 ? \
1686                 ippiFilterGaussianBorder_##ippfavor##_C1R(src.ptr<ippType>(), (int)src.step, \
1687                     dst.ptr<ippType>(), (int)dst.step, roiSize, borderValues[0], pSpec, pBuffer) : \
1688                 ippiFilterGaussianBorder_##ippfavor##_C3R(src.ptr<ippType>(), (int)src.step, \
1689                     dst.ptr<ippType>(), (int)dst.step, roiSize, borderValues, pSpec, pBuffer); \
1690             ippFree(pBuffer); \
1691             ippFree(pSpec); \
1692             if (status >= 0) \
1693             { \
1694                 CV_IMPL_ADD(CV_IMPL_IPP); \
1695                 return; \
1696             } \
1697         } while ((void)0, 0)
1698 
1699                         if (type == CV_8UC1)
1700                             IPP_FILTER_GAUSS(8u, 1);
1701                         else if (type == CV_8UC3)
1702                             IPP_FILTER_GAUSS(8u, 3);
1703                         else if (type == CV_16UC1)
1704                             IPP_FILTER_GAUSS(16u, 1);
1705                         else if (type == CV_16UC3)
1706                             IPP_FILTER_GAUSS(16u, 3);
1707                         else if (type == CV_16SC1)
1708                             IPP_FILTER_GAUSS(16s, 1);
1709                         else if (type == CV_16SC3)
1710                             IPP_FILTER_GAUSS(16s, 3);
1711                         else if (type == CV_32FC1)
1712                             IPP_FILTER_GAUSS(32f, 1);
1713                         else if (type == CV_32FC3)
1714                             IPP_FILTER_GAUSS(32f, 3);
1715 #undef IPP_FILTER_GAUSS
1716                     }
1717                 }
1718                 setIppErrorStatus();
1719             }
1720         }
1721     }
1722 #endif
1723 
1724     Mat kx, ky;
1725     createGaussianKernels(kx, ky, type, ksize, sigma1, sigma2);
1726     sepFilter2D(_src, _dst, CV_MAT_DEPTH(type), kx, ky, Point(-1,-1), 0, borderType );
1727 }
1728 
1729 /****************************************************************************************\
1730                                       Median Filter
1731 \****************************************************************************************/
1732 
1733 namespace cv
1734 {
1735 typedef ushort HT;
1736 
1737 /**
1738  * This structure represents a two-tier histogram. The first tier (known as the
1739  * "coarse" level) is 4 bit wide and the second tier (known as the "fine" level)
1740  * is 8 bit wide. Pixels inserted in the fine level also get inserted into the
1741  * coarse bucket designated by the 4 MSBs of the fine bucket value.
1742  *
1743  * The structure is aligned on 16 bits, which is a prerequisite for SIMD
1744  * instructions. Each bucket is 16 bit wide, which means that extra care must be
1745  * taken to prevent overflow.
1746  */
1747 typedef struct
1748 {
1749     HT coarse[16];
1750     HT fine[16][16];
1751 } Histogram;
1752 
1753 
1754 #if CV_SSE2
1755 #define MEDIAN_HAVE_SIMD 1
1756 
histogram_add_simd(const HT x[16],HT y[16])1757 static inline void histogram_add_simd( const HT x[16], HT y[16] )
1758 {
1759     const __m128i* rx = (const __m128i*)x;
1760     __m128i* ry = (__m128i*)y;
1761     __m128i r0 = _mm_add_epi16(_mm_load_si128(ry+0),_mm_load_si128(rx+0));
1762     __m128i r1 = _mm_add_epi16(_mm_load_si128(ry+1),_mm_load_si128(rx+1));
1763     _mm_store_si128(ry+0, r0);
1764     _mm_store_si128(ry+1, r1);
1765 }
1766 
histogram_sub_simd(const HT x[16],HT y[16])1767 static inline void histogram_sub_simd( const HT x[16], HT y[16] )
1768 {
1769     const __m128i* rx = (const __m128i*)x;
1770     __m128i* ry = (__m128i*)y;
1771     __m128i r0 = _mm_sub_epi16(_mm_load_si128(ry+0),_mm_load_si128(rx+0));
1772     __m128i r1 = _mm_sub_epi16(_mm_load_si128(ry+1),_mm_load_si128(rx+1));
1773     _mm_store_si128(ry+0, r0);
1774     _mm_store_si128(ry+1, r1);
1775 }
1776 
1777 #elif CV_NEON
1778 #define MEDIAN_HAVE_SIMD 1
1779 
histogram_add_simd(const HT x[16],HT y[16])1780 static inline void histogram_add_simd( const HT x[16], HT y[16] )
1781 {
1782     vst1q_u16(y, vaddq_u16(vld1q_u16(x), vld1q_u16(y)));
1783     vst1q_u16(y + 8, vaddq_u16(vld1q_u16(x + 8), vld1q_u16(y + 8)));
1784 }
1785 
histogram_sub_simd(const HT x[16],HT y[16])1786 static inline void histogram_sub_simd( const HT x[16], HT y[16] )
1787 {
1788     vst1q_u16(y, vsubq_u16(vld1q_u16(x), vld1q_u16(y)));
1789     vst1q_u16(y + 8, vsubq_u16(vld1q_u16(x + 8), vld1q_u16(y + 8)));
1790 }
1791 
1792 #else
1793 #define MEDIAN_HAVE_SIMD 0
1794 #endif
1795 
1796 
histogram_add(const HT x[16],HT y[16])1797 static inline void histogram_add( const HT x[16], HT y[16] )
1798 {
1799     int i;
1800     for( i = 0; i < 16; ++i )
1801         y[i] = (HT)(y[i] + x[i]);
1802 }
1803 
histogram_sub(const HT x[16],HT y[16])1804 static inline void histogram_sub( const HT x[16], HT y[16] )
1805 {
1806     int i;
1807     for( i = 0; i < 16; ++i )
1808         y[i] = (HT)(y[i] - x[i]);
1809 }
1810 
histogram_muladd(int a,const HT x[16],HT y[16])1811 static inline void histogram_muladd( int a, const HT x[16],
1812         HT y[16] )
1813 {
1814     for( int i = 0; i < 16; ++i )
1815         y[i] = (HT)(y[i] + a * x[i]);
1816 }
1817 
1818 static void
medianBlur_8u_O1(const Mat & _src,Mat & _dst,int ksize)1819 medianBlur_8u_O1( const Mat& _src, Mat& _dst, int ksize )
1820 {
1821 /**
1822  * HOP is short for Histogram OPeration. This macro makes an operation \a op on
1823  * histogram \a h for pixel value \a x. It takes care of handling both levels.
1824  */
1825 #define HOP(h,x,op) \
1826     h.coarse[x>>4] op, \
1827     *((HT*)h.fine + x) op
1828 
1829 #define COP(c,j,x,op) \
1830     h_coarse[ 16*(n*c+j) + (x>>4) ] op, \
1831     h_fine[ 16 * (n*(16*c+(x>>4)) + j) + (x & 0xF) ] op
1832 
1833     int cn = _dst.channels(), m = _dst.rows, r = (ksize-1)/2;
1834     size_t sstep = _src.step, dstep = _dst.step;
1835     Histogram CV_DECL_ALIGNED(16) H[4];
1836     HT CV_DECL_ALIGNED(16) luc[4][16];
1837 
1838     int STRIPE_SIZE = std::min( _dst.cols, 512/cn );
1839 
1840     std::vector<HT> _h_coarse(1 * 16 * (STRIPE_SIZE + 2*r) * cn + 16);
1841     std::vector<HT> _h_fine(16 * 16 * (STRIPE_SIZE + 2*r) * cn + 16);
1842     HT* h_coarse = alignPtr(&_h_coarse[0], 16);
1843     HT* h_fine = alignPtr(&_h_fine[0], 16);
1844 #if MEDIAN_HAVE_SIMD
1845     volatile bool useSIMD = checkHardwareSupport(CV_CPU_SSE2) || checkHardwareSupport(CV_CPU_NEON);
1846 #endif
1847 
1848     for( int x = 0; x < _dst.cols; x += STRIPE_SIZE )
1849     {
1850         int i, j, k, c, n = std::min(_dst.cols - x, STRIPE_SIZE) + r*2;
1851         const uchar* src = _src.ptr() + x*cn;
1852         uchar* dst = _dst.ptr() + (x - r)*cn;
1853 
1854         memset( h_coarse, 0, 16*n*cn*sizeof(h_coarse[0]) );
1855         memset( h_fine, 0, 16*16*n*cn*sizeof(h_fine[0]) );
1856 
1857         // First row initialization
1858         for( c = 0; c < cn; c++ )
1859         {
1860             for( j = 0; j < n; j++ )
1861                 COP( c, j, src[cn*j+c], += (cv::HT)(r+2) );
1862 
1863             for( i = 1; i < r; i++ )
1864             {
1865                 const uchar* p = src + sstep*std::min(i, m-1);
1866                 for ( j = 0; j < n; j++ )
1867                     COP( c, j, p[cn*j+c], ++ );
1868             }
1869         }
1870 
1871         for( i = 0; i < m; i++ )
1872         {
1873             const uchar* p0 = src + sstep * std::max( 0, i-r-1 );
1874             const uchar* p1 = src + sstep * std::min( m-1, i+r );
1875 
1876             memset( H, 0, cn*sizeof(H[0]) );
1877             memset( luc, 0, cn*sizeof(luc[0]) );
1878             for( c = 0; c < cn; c++ )
1879             {
1880                 // Update column histograms for the entire row.
1881                 for( j = 0; j < n; j++ )
1882                 {
1883                     COP( c, j, p0[j*cn + c], -- );
1884                     COP( c, j, p1[j*cn + c], ++ );
1885                 }
1886 
1887                 // First column initialization
1888                 for( k = 0; k < 16; ++k )
1889                     histogram_muladd( 2*r+1, &h_fine[16*n*(16*c+k)], &H[c].fine[k][0] );
1890 
1891             #if MEDIAN_HAVE_SIMD
1892                 if( useSIMD )
1893                 {
1894                     for( j = 0; j < 2*r; ++j )
1895                         histogram_add_simd( &h_coarse[16*(n*c+j)], H[c].coarse );
1896 
1897                     for( j = r; j < n-r; j++ )
1898                     {
1899                         int t = 2*r*r + 2*r, b, sum = 0;
1900                         HT* segment;
1901 
1902                         histogram_add_simd( &h_coarse[16*(n*c + std::min(j+r,n-1))], H[c].coarse );
1903 
1904                         // Find median at coarse level
1905                         for ( k = 0; k < 16 ; ++k )
1906                         {
1907                             sum += H[c].coarse[k];
1908                             if ( sum > t )
1909                             {
1910                                 sum -= H[c].coarse[k];
1911                                 break;
1912                             }
1913                         }
1914                         assert( k < 16 );
1915 
1916                         /* Update corresponding histogram segment */
1917                         if ( luc[c][k] <= j-r )
1918                         {
1919                             memset( &H[c].fine[k], 0, 16 * sizeof(HT) );
1920                             for ( luc[c][k] = cv::HT(j-r); luc[c][k] < MIN(j+r+1,n); ++luc[c][k] )
1921                                 histogram_add_simd( &h_fine[16*(n*(16*c+k)+luc[c][k])], H[c].fine[k] );
1922 
1923                             if ( luc[c][k] < j+r+1 )
1924                             {
1925                                 histogram_muladd( j+r+1 - n, &h_fine[16*(n*(16*c+k)+(n-1))], &H[c].fine[k][0] );
1926                                 luc[c][k] = (HT)(j+r+1);
1927                             }
1928                         }
1929                         else
1930                         {
1931                             for ( ; luc[c][k] < j+r+1; ++luc[c][k] )
1932                             {
1933                                 histogram_sub_simd( &h_fine[16*(n*(16*c+k)+MAX(luc[c][k]-2*r-1,0))], H[c].fine[k] );
1934                                 histogram_add_simd( &h_fine[16*(n*(16*c+k)+MIN(luc[c][k],n-1))], H[c].fine[k] );
1935                             }
1936                         }
1937 
1938                         histogram_sub_simd( &h_coarse[16*(n*c+MAX(j-r,0))], H[c].coarse );
1939 
1940                         /* Find median in segment */
1941                         segment = H[c].fine[k];
1942                         for ( b = 0; b < 16 ; b++ )
1943                         {
1944                             sum += segment[b];
1945                             if ( sum > t )
1946                             {
1947                                 dst[dstep*i+cn*j+c] = (uchar)(16*k + b);
1948                                 break;
1949                             }
1950                         }
1951                         assert( b < 16 );
1952                     }
1953                 }
1954                 else
1955             #endif
1956                 {
1957                     for( j = 0; j < 2*r; ++j )
1958                         histogram_add( &h_coarse[16*(n*c+j)], H[c].coarse );
1959 
1960                     for( j = r; j < n-r; j++ )
1961                     {
1962                         int t = 2*r*r + 2*r, b, sum = 0;
1963                         HT* segment;
1964 
1965                         histogram_add( &h_coarse[16*(n*c + std::min(j+r,n-1))], H[c].coarse );
1966 
1967                         // Find median at coarse level
1968                         for ( k = 0; k < 16 ; ++k )
1969                         {
1970                             sum += H[c].coarse[k];
1971                             if ( sum > t )
1972                             {
1973                                 sum -= H[c].coarse[k];
1974                                 break;
1975                             }
1976                         }
1977                         assert( k < 16 );
1978 
1979                         /* Update corresponding histogram segment */
1980                         if ( luc[c][k] <= j-r )
1981                         {
1982                             memset( &H[c].fine[k], 0, 16 * sizeof(HT) );
1983                             for ( luc[c][k] = cv::HT(j-r); luc[c][k] < MIN(j+r+1,n); ++luc[c][k] )
1984                                 histogram_add( &h_fine[16*(n*(16*c+k)+luc[c][k])], H[c].fine[k] );
1985 
1986                             if ( luc[c][k] < j+r+1 )
1987                             {
1988                                 histogram_muladd( j+r+1 - n, &h_fine[16*(n*(16*c+k)+(n-1))], &H[c].fine[k][0] );
1989                                 luc[c][k] = (HT)(j+r+1);
1990                             }
1991                         }
1992                         else
1993                         {
1994                             for ( ; luc[c][k] < j+r+1; ++luc[c][k] )
1995                             {
1996                                 histogram_sub( &h_fine[16*(n*(16*c+k)+MAX(luc[c][k]-2*r-1,0))], H[c].fine[k] );
1997                                 histogram_add( &h_fine[16*(n*(16*c+k)+MIN(luc[c][k],n-1))], H[c].fine[k] );
1998                             }
1999                         }
2000 
2001                         histogram_sub( &h_coarse[16*(n*c+MAX(j-r,0))], H[c].coarse );
2002 
2003                         /* Find median in segment */
2004                         segment = H[c].fine[k];
2005                         for ( b = 0; b < 16 ; b++ )
2006                         {
2007                             sum += segment[b];
2008                             if ( sum > t )
2009                             {
2010                                 dst[dstep*i+cn*j+c] = (uchar)(16*k + b);
2011                                 break;
2012                             }
2013                         }
2014                         assert( b < 16 );
2015                     }
2016                 }
2017             }
2018         }
2019     }
2020 
2021 #undef HOP
2022 #undef COP
2023 }
2024 
2025 static void
medianBlur_8u_Om(const Mat & _src,Mat & _dst,int m)2026 medianBlur_8u_Om( const Mat& _src, Mat& _dst, int m )
2027 {
2028     #define N  16
2029     int     zone0[4][N];
2030     int     zone1[4][N*N];
2031     int     x, y;
2032     int     n2 = m*m/2;
2033     Size    size = _dst.size();
2034     const uchar* src = _src.ptr();
2035     uchar*  dst = _dst.ptr();
2036     int     src_step = (int)_src.step, dst_step = (int)_dst.step;
2037     int     cn = _src.channels();
2038     const uchar*  src_max = src + size.height*src_step;
2039 
2040     #define UPDATE_ACC01( pix, cn, op ) \
2041     {                                   \
2042         int p = (pix);                  \
2043         zone1[cn][p] op;                \
2044         zone0[cn][p >> 4] op;           \
2045     }
2046 
2047     //CV_Assert( size.height >= nx && size.width >= nx );
2048     for( x = 0; x < size.width; x++, src += cn, dst += cn )
2049     {
2050         uchar* dst_cur = dst;
2051         const uchar* src_top = src;
2052         const uchar* src_bottom = src;
2053         int k, c;
2054         int src_step1 = src_step, dst_step1 = dst_step;
2055 
2056         if( x % 2 != 0 )
2057         {
2058             src_bottom = src_top += src_step*(size.height-1);
2059             dst_cur += dst_step*(size.height-1);
2060             src_step1 = -src_step1;
2061             dst_step1 = -dst_step1;
2062         }
2063 
2064         // init accumulator
2065         memset( zone0, 0, sizeof(zone0[0])*cn );
2066         memset( zone1, 0, sizeof(zone1[0])*cn );
2067 
2068         for( y = 0; y <= m/2; y++ )
2069         {
2070             for( c = 0; c < cn; c++ )
2071             {
2072                 if( y > 0 )
2073                 {
2074                     for( k = 0; k < m*cn; k += cn )
2075                         UPDATE_ACC01( src_bottom[k+c], c, ++ );
2076                 }
2077                 else
2078                 {
2079                     for( k = 0; k < m*cn; k += cn )
2080                         UPDATE_ACC01( src_bottom[k+c], c, += m/2+1 );
2081                 }
2082             }
2083 
2084             if( (src_step1 > 0 && y < size.height-1) ||
2085                 (src_step1 < 0 && size.height-y-1 > 0) )
2086                 src_bottom += src_step1;
2087         }
2088 
2089         for( y = 0; y < size.height; y++, dst_cur += dst_step1 )
2090         {
2091             // find median
2092             for( c = 0; c < cn; c++ )
2093             {
2094                 int s = 0;
2095                 for( k = 0; ; k++ )
2096                 {
2097                     int t = s + zone0[c][k];
2098                     if( t > n2 ) break;
2099                     s = t;
2100                 }
2101 
2102                 for( k *= N; ;k++ )
2103                 {
2104                     s += zone1[c][k];
2105                     if( s > n2 ) break;
2106                 }
2107 
2108                 dst_cur[c] = (uchar)k;
2109             }
2110 
2111             if( y+1 == size.height )
2112                 break;
2113 
2114             if( cn == 1 )
2115             {
2116                 for( k = 0; k < m; k++ )
2117                 {
2118                     int p = src_top[k];
2119                     int q = src_bottom[k];
2120                     zone1[0][p]--;
2121                     zone0[0][p>>4]--;
2122                     zone1[0][q]++;
2123                     zone0[0][q>>4]++;
2124                 }
2125             }
2126             else if( cn == 3 )
2127             {
2128                 for( k = 0; k < m*3; k += 3 )
2129                 {
2130                     UPDATE_ACC01( src_top[k], 0, -- );
2131                     UPDATE_ACC01( src_top[k+1], 1, -- );
2132                     UPDATE_ACC01( src_top[k+2], 2, -- );
2133 
2134                     UPDATE_ACC01( src_bottom[k], 0, ++ );
2135                     UPDATE_ACC01( src_bottom[k+1], 1, ++ );
2136                     UPDATE_ACC01( src_bottom[k+2], 2, ++ );
2137                 }
2138             }
2139             else
2140             {
2141                 assert( cn == 4 );
2142                 for( k = 0; k < m*4; k += 4 )
2143                 {
2144                     UPDATE_ACC01( src_top[k], 0, -- );
2145                     UPDATE_ACC01( src_top[k+1], 1, -- );
2146                     UPDATE_ACC01( src_top[k+2], 2, -- );
2147                     UPDATE_ACC01( src_top[k+3], 3, -- );
2148 
2149                     UPDATE_ACC01( src_bottom[k], 0, ++ );
2150                     UPDATE_ACC01( src_bottom[k+1], 1, ++ );
2151                     UPDATE_ACC01( src_bottom[k+2], 2, ++ );
2152                     UPDATE_ACC01( src_bottom[k+3], 3, ++ );
2153                 }
2154             }
2155 
2156             if( (src_step1 > 0 && src_bottom + src_step1 < src_max) ||
2157                 (src_step1 < 0 && src_bottom + src_step1 >= src) )
2158                 src_bottom += src_step1;
2159 
2160             if( y >= m/2 )
2161                 src_top += src_step1;
2162         }
2163     }
2164 #undef N
2165 #undef UPDATE_ACC
2166 }
2167 
2168 
2169 struct MinMax8u
2170 {
2171     typedef uchar value_type;
2172     typedef int arg_type;
2173     enum { SIZE = 1 };
loadcv::MinMax8u2174     arg_type load(const uchar* ptr) { return *ptr; }
storecv::MinMax8u2175     void store(uchar* ptr, arg_type val) { *ptr = (uchar)val; }
operator ()cv::MinMax8u2176     void operator()(arg_type& a, arg_type& b) const
2177     {
2178         int t = CV_FAST_CAST_8U(a - b);
2179         b += t; a -= t;
2180     }
2181 };
2182 
2183 struct MinMax16u
2184 {
2185     typedef ushort value_type;
2186     typedef int arg_type;
2187     enum { SIZE = 1 };
loadcv::MinMax16u2188     arg_type load(const ushort* ptr) { return *ptr; }
storecv::MinMax16u2189     void store(ushort* ptr, arg_type val) { *ptr = (ushort)val; }
operator ()cv::MinMax16u2190     void operator()(arg_type& a, arg_type& b) const
2191     {
2192         arg_type t = a;
2193         a = std::min(a, b);
2194         b = std::max(b, t);
2195     }
2196 };
2197 
2198 struct MinMax16s
2199 {
2200     typedef short value_type;
2201     typedef int arg_type;
2202     enum { SIZE = 1 };
loadcv::MinMax16s2203     arg_type load(const short* ptr) { return *ptr; }
storecv::MinMax16s2204     void store(short* ptr, arg_type val) { *ptr = (short)val; }
operator ()cv::MinMax16s2205     void operator()(arg_type& a, arg_type& b) const
2206     {
2207         arg_type t = a;
2208         a = std::min(a, b);
2209         b = std::max(b, t);
2210     }
2211 };
2212 
2213 struct MinMax32f
2214 {
2215     typedef float value_type;
2216     typedef float arg_type;
2217     enum { SIZE = 1 };
loadcv::MinMax32f2218     arg_type load(const float* ptr) { return *ptr; }
storecv::MinMax32f2219     void store(float* ptr, arg_type val) { *ptr = val; }
operator ()cv::MinMax32f2220     void operator()(arg_type& a, arg_type& b) const
2221     {
2222         arg_type t = a;
2223         a = std::min(a, b);
2224         b = std::max(b, t);
2225     }
2226 };
2227 
2228 #if CV_SSE2
2229 
2230 struct MinMaxVec8u
2231 {
2232     typedef uchar value_type;
2233     typedef __m128i arg_type;
2234     enum { SIZE = 16 };
loadcv::MinMaxVec8u2235     arg_type load(const uchar* ptr) { return _mm_loadu_si128((const __m128i*)ptr); }
storecv::MinMaxVec8u2236     void store(uchar* ptr, arg_type val) { _mm_storeu_si128((__m128i*)ptr, val); }
operator ()cv::MinMaxVec8u2237     void operator()(arg_type& a, arg_type& b) const
2238     {
2239         arg_type t = a;
2240         a = _mm_min_epu8(a, b);
2241         b = _mm_max_epu8(b, t);
2242     }
2243 };
2244 
2245 
2246 struct MinMaxVec16u
2247 {
2248     typedef ushort value_type;
2249     typedef __m128i arg_type;
2250     enum { SIZE = 8 };
loadcv::MinMaxVec16u2251     arg_type load(const ushort* ptr) { return _mm_loadu_si128((const __m128i*)ptr); }
storecv::MinMaxVec16u2252     void store(ushort* ptr, arg_type val) { _mm_storeu_si128((__m128i*)ptr, val); }
operator ()cv::MinMaxVec16u2253     void operator()(arg_type& a, arg_type& b) const
2254     {
2255         arg_type t = _mm_subs_epu16(a, b);
2256         a = _mm_subs_epu16(a, t);
2257         b = _mm_adds_epu16(b, t);
2258     }
2259 };
2260 
2261 
2262 struct MinMaxVec16s
2263 {
2264     typedef short value_type;
2265     typedef __m128i arg_type;
2266     enum { SIZE = 8 };
loadcv::MinMaxVec16s2267     arg_type load(const short* ptr) { return _mm_loadu_si128((const __m128i*)ptr); }
storecv::MinMaxVec16s2268     void store(short* ptr, arg_type val) { _mm_storeu_si128((__m128i*)ptr, val); }
operator ()cv::MinMaxVec16s2269     void operator()(arg_type& a, arg_type& b) const
2270     {
2271         arg_type t = a;
2272         a = _mm_min_epi16(a, b);
2273         b = _mm_max_epi16(b, t);
2274     }
2275 };
2276 
2277 
2278 struct MinMaxVec32f
2279 {
2280     typedef float value_type;
2281     typedef __m128 arg_type;
2282     enum { SIZE = 4 };
loadcv::MinMaxVec32f2283     arg_type load(const float* ptr) { return _mm_loadu_ps(ptr); }
storecv::MinMaxVec32f2284     void store(float* ptr, arg_type val) { _mm_storeu_ps(ptr, val); }
operator ()cv::MinMaxVec32f2285     void operator()(arg_type& a, arg_type& b) const
2286     {
2287         arg_type t = a;
2288         a = _mm_min_ps(a, b);
2289         b = _mm_max_ps(b, t);
2290     }
2291 };
2292 
2293 #elif CV_NEON
2294 
2295 struct MinMaxVec8u
2296 {
2297     typedef uchar value_type;
2298     typedef uint8x16_t arg_type;
2299     enum { SIZE = 16 };
loadcv::MinMaxVec8u2300     arg_type load(const uchar* ptr) { return vld1q_u8(ptr); }
storecv::MinMaxVec8u2301     void store(uchar* ptr, arg_type val) { vst1q_u8(ptr, val); }
operator ()cv::MinMaxVec8u2302     void operator()(arg_type& a, arg_type& b) const
2303     {
2304         arg_type t = a;
2305         a = vminq_u8(a, b);
2306         b = vmaxq_u8(b, t);
2307     }
2308 };
2309 
2310 
2311 struct MinMaxVec16u
2312 {
2313     typedef ushort value_type;
2314     typedef uint16x8_t arg_type;
2315     enum { SIZE = 8 };
loadcv::MinMaxVec16u2316     arg_type load(const ushort* ptr) { return vld1q_u16(ptr); }
storecv::MinMaxVec16u2317     void store(ushort* ptr, arg_type val) { vst1q_u16(ptr, val); }
operator ()cv::MinMaxVec16u2318     void operator()(arg_type& a, arg_type& b) const
2319     {
2320         arg_type t = a;
2321         a = vminq_u16(a, b);
2322         b = vmaxq_u16(b, t);
2323     }
2324 };
2325 
2326 
2327 struct MinMaxVec16s
2328 {
2329     typedef short value_type;
2330     typedef int16x8_t arg_type;
2331     enum { SIZE = 8 };
loadcv::MinMaxVec16s2332     arg_type load(const short* ptr) { return vld1q_s16(ptr); }
storecv::MinMaxVec16s2333     void store(short* ptr, arg_type val) { vst1q_s16(ptr, val); }
operator ()cv::MinMaxVec16s2334     void operator()(arg_type& a, arg_type& b) const
2335     {
2336         arg_type t = a;
2337         a = vminq_s16(a, b);
2338         b = vmaxq_s16(b, t);
2339     }
2340 };
2341 
2342 
2343 struct MinMaxVec32f
2344 {
2345     typedef float value_type;
2346     typedef float32x4_t arg_type;
2347     enum { SIZE = 4 };
loadcv::MinMaxVec32f2348     arg_type load(const float* ptr) { return vld1q_f32(ptr); }
storecv::MinMaxVec32f2349     void store(float* ptr, arg_type val) { vst1q_f32(ptr, val); }
operator ()cv::MinMaxVec32f2350     void operator()(arg_type& a, arg_type& b) const
2351     {
2352         arg_type t = a;
2353         a = vminq_f32(a, b);
2354         b = vmaxq_f32(b, t);
2355     }
2356 };
2357 
2358 
2359 #else
2360 
2361 typedef MinMax8u MinMaxVec8u;
2362 typedef MinMax16u MinMaxVec16u;
2363 typedef MinMax16s MinMaxVec16s;
2364 typedef MinMax32f MinMaxVec32f;
2365 
2366 #endif
2367 
2368 template<class Op, class VecOp>
2369 static void
medianBlur_SortNet(const Mat & _src,Mat & _dst,int m)2370 medianBlur_SortNet( const Mat& _src, Mat& _dst, int m )
2371 {
2372     typedef typename Op::value_type T;
2373     typedef typename Op::arg_type WT;
2374     typedef typename VecOp::arg_type VT;
2375 
2376     const T* src = _src.ptr<T>();
2377     T* dst = _dst.ptr<T>();
2378     int sstep = (int)(_src.step/sizeof(T));
2379     int dstep = (int)(_dst.step/sizeof(T));
2380     Size size = _dst.size();
2381     int i, j, k, cn = _src.channels();
2382     Op op;
2383     VecOp vop;
2384     volatile bool useSIMD = checkHardwareSupport(CV_CPU_SSE2) || checkHardwareSupport(CV_CPU_NEON);
2385 
2386     if( m == 3 )
2387     {
2388         if( size.width == 1 || size.height == 1 )
2389         {
2390             int len = size.width + size.height - 1;
2391             int sdelta = size.height == 1 ? cn : sstep;
2392             int sdelta0 = size.height == 1 ? 0 : sstep - cn;
2393             int ddelta = size.height == 1 ? cn : dstep;
2394 
2395             for( i = 0; i < len; i++, src += sdelta0, dst += ddelta )
2396                 for( j = 0; j < cn; j++, src++ )
2397                 {
2398                     WT p0 = src[i > 0 ? -sdelta : 0];
2399                     WT p1 = src[0];
2400                     WT p2 = src[i < len - 1 ? sdelta : 0];
2401 
2402                     op(p0, p1); op(p1, p2); op(p0, p1);
2403                     dst[j] = (T)p1;
2404                 }
2405             return;
2406         }
2407 
2408         size.width *= cn;
2409         for( i = 0; i < size.height; i++, dst += dstep )
2410         {
2411             const T* row0 = src + std::max(i - 1, 0)*sstep;
2412             const T* row1 = src + i*sstep;
2413             const T* row2 = src + std::min(i + 1, size.height-1)*sstep;
2414             int limit = useSIMD ? cn : size.width;
2415 
2416             for(j = 0;; )
2417             {
2418                 for( ; j < limit; j++ )
2419                 {
2420                     int j0 = j >= cn ? j - cn : j;
2421                     int j2 = j < size.width - cn ? j + cn : j;
2422                     WT p0 = row0[j0], p1 = row0[j], p2 = row0[j2];
2423                     WT p3 = row1[j0], p4 = row1[j], p5 = row1[j2];
2424                     WT p6 = row2[j0], p7 = row2[j], p8 = row2[j2];
2425 
2426                     op(p1, p2); op(p4, p5); op(p7, p8); op(p0, p1);
2427                     op(p3, p4); op(p6, p7); op(p1, p2); op(p4, p5);
2428                     op(p7, p8); op(p0, p3); op(p5, p8); op(p4, p7);
2429                     op(p3, p6); op(p1, p4); op(p2, p5); op(p4, p7);
2430                     op(p4, p2); op(p6, p4); op(p4, p2);
2431                     dst[j] = (T)p4;
2432                 }
2433 
2434                 if( limit == size.width )
2435                     break;
2436 
2437                 for( ; j <= size.width - VecOp::SIZE - cn; j += VecOp::SIZE )
2438                 {
2439                     VT p0 = vop.load(row0+j-cn), p1 = vop.load(row0+j), p2 = vop.load(row0+j+cn);
2440                     VT p3 = vop.load(row1+j-cn), p4 = vop.load(row1+j), p5 = vop.load(row1+j+cn);
2441                     VT p6 = vop.load(row2+j-cn), p7 = vop.load(row2+j), p8 = vop.load(row2+j+cn);
2442 
2443                     vop(p1, p2); vop(p4, p5); vop(p7, p8); vop(p0, p1);
2444                     vop(p3, p4); vop(p6, p7); vop(p1, p2); vop(p4, p5);
2445                     vop(p7, p8); vop(p0, p3); vop(p5, p8); vop(p4, p7);
2446                     vop(p3, p6); vop(p1, p4); vop(p2, p5); vop(p4, p7);
2447                     vop(p4, p2); vop(p6, p4); vop(p4, p2);
2448                     vop.store(dst+j, p4);
2449                 }
2450 
2451                 limit = size.width;
2452             }
2453         }
2454     }
2455     else if( m == 5 )
2456     {
2457         if( size.width == 1 || size.height == 1 )
2458         {
2459             int len = size.width + size.height - 1;
2460             int sdelta = size.height == 1 ? cn : sstep;
2461             int sdelta0 = size.height == 1 ? 0 : sstep - cn;
2462             int ddelta = size.height == 1 ? cn : dstep;
2463 
2464             for( i = 0; i < len; i++, src += sdelta0, dst += ddelta )
2465                 for( j = 0; j < cn; j++, src++ )
2466                 {
2467                     int i1 = i > 0 ? -sdelta : 0;
2468                     int i0 = i > 1 ? -sdelta*2 : i1;
2469                     int i3 = i < len-1 ? sdelta : 0;
2470                     int i4 = i < len-2 ? sdelta*2 : i3;
2471                     WT p0 = src[i0], p1 = src[i1], p2 = src[0], p3 = src[i3], p4 = src[i4];
2472 
2473                     op(p0, p1); op(p3, p4); op(p2, p3); op(p3, p4); op(p0, p2);
2474                     op(p2, p4); op(p1, p3); op(p1, p2);
2475                     dst[j] = (T)p2;
2476                 }
2477             return;
2478         }
2479 
2480         size.width *= cn;
2481         for( i = 0; i < size.height; i++, dst += dstep )
2482         {
2483             const T* row[5];
2484             row[0] = src + std::max(i - 2, 0)*sstep;
2485             row[1] = src + std::max(i - 1, 0)*sstep;
2486             row[2] = src + i*sstep;
2487             row[3] = src + std::min(i + 1, size.height-1)*sstep;
2488             row[4] = src + std::min(i + 2, size.height-1)*sstep;
2489             int limit = useSIMD ? cn*2 : size.width;
2490 
2491             for(j = 0;; )
2492             {
2493                 for( ; j < limit; j++ )
2494                 {
2495                     WT p[25];
2496                     int j1 = j >= cn ? j - cn : j;
2497                     int j0 = j >= cn*2 ? j - cn*2 : j1;
2498                     int j3 = j < size.width - cn ? j + cn : j;
2499                     int j4 = j < size.width - cn*2 ? j + cn*2 : j3;
2500                     for( k = 0; k < 5; k++ )
2501                     {
2502                         const T* rowk = row[k];
2503                         p[k*5] = rowk[j0]; p[k*5+1] = rowk[j1];
2504                         p[k*5+2] = rowk[j]; p[k*5+3] = rowk[j3];
2505                         p[k*5+4] = rowk[j4];
2506                     }
2507 
2508                     op(p[1], p[2]); op(p[0], p[1]); op(p[1], p[2]); op(p[4], p[5]); op(p[3], p[4]);
2509                     op(p[4], p[5]); op(p[0], p[3]); op(p[2], p[5]); op(p[2], p[3]); op(p[1], p[4]);
2510                     op(p[1], p[2]); op(p[3], p[4]); op(p[7], p[8]); op(p[6], p[7]); op(p[7], p[8]);
2511                     op(p[10], p[11]); op(p[9], p[10]); op(p[10], p[11]); op(p[6], p[9]); op(p[8], p[11]);
2512                     op(p[8], p[9]); op(p[7], p[10]); op(p[7], p[8]); op(p[9], p[10]); op(p[0], p[6]);
2513                     op(p[4], p[10]); op(p[4], p[6]); op(p[2], p[8]); op(p[2], p[4]); op(p[6], p[8]);
2514                     op(p[1], p[7]); op(p[5], p[11]); op(p[5], p[7]); op(p[3], p[9]); op(p[3], p[5]);
2515                     op(p[7], p[9]); op(p[1], p[2]); op(p[3], p[4]); op(p[5], p[6]); op(p[7], p[8]);
2516                     op(p[9], p[10]); op(p[13], p[14]); op(p[12], p[13]); op(p[13], p[14]); op(p[16], p[17]);
2517                     op(p[15], p[16]); op(p[16], p[17]); op(p[12], p[15]); op(p[14], p[17]); op(p[14], p[15]);
2518                     op(p[13], p[16]); op(p[13], p[14]); op(p[15], p[16]); op(p[19], p[20]); op(p[18], p[19]);
2519                     op(p[19], p[20]); op(p[21], p[22]); op(p[23], p[24]); op(p[21], p[23]); op(p[22], p[24]);
2520                     op(p[22], p[23]); op(p[18], p[21]); op(p[20], p[23]); op(p[20], p[21]); op(p[19], p[22]);
2521                     op(p[22], p[24]); op(p[19], p[20]); op(p[21], p[22]); op(p[23], p[24]); op(p[12], p[18]);
2522                     op(p[16], p[22]); op(p[16], p[18]); op(p[14], p[20]); op(p[20], p[24]); op(p[14], p[16]);
2523                     op(p[18], p[20]); op(p[22], p[24]); op(p[13], p[19]); op(p[17], p[23]); op(p[17], p[19]);
2524                     op(p[15], p[21]); op(p[15], p[17]); op(p[19], p[21]); op(p[13], p[14]); op(p[15], p[16]);
2525                     op(p[17], p[18]); op(p[19], p[20]); op(p[21], p[22]); op(p[23], p[24]); op(p[0], p[12]);
2526                     op(p[8], p[20]); op(p[8], p[12]); op(p[4], p[16]); op(p[16], p[24]); op(p[12], p[16]);
2527                     op(p[2], p[14]); op(p[10], p[22]); op(p[10], p[14]); op(p[6], p[18]); op(p[6], p[10]);
2528                     op(p[10], p[12]); op(p[1], p[13]); op(p[9], p[21]); op(p[9], p[13]); op(p[5], p[17]);
2529                     op(p[13], p[17]); op(p[3], p[15]); op(p[11], p[23]); op(p[11], p[15]); op(p[7], p[19]);
2530                     op(p[7], p[11]); op(p[11], p[13]); op(p[11], p[12]);
2531                     dst[j] = (T)p[12];
2532                 }
2533 
2534                 if( limit == size.width )
2535                     break;
2536 
2537                 for( ; j <= size.width - VecOp::SIZE - cn*2; j += VecOp::SIZE )
2538                 {
2539                     VT p[25];
2540                     for( k = 0; k < 5; k++ )
2541                     {
2542                         const T* rowk = row[k];
2543                         p[k*5] = vop.load(rowk+j-cn*2); p[k*5+1] = vop.load(rowk+j-cn);
2544                         p[k*5+2] = vop.load(rowk+j); p[k*5+3] = vop.load(rowk+j+cn);
2545                         p[k*5+4] = vop.load(rowk+j+cn*2);
2546                     }
2547 
2548                     vop(p[1], p[2]); vop(p[0], p[1]); vop(p[1], p[2]); vop(p[4], p[5]); vop(p[3], p[4]);
2549                     vop(p[4], p[5]); vop(p[0], p[3]); vop(p[2], p[5]); vop(p[2], p[3]); vop(p[1], p[4]);
2550                     vop(p[1], p[2]); vop(p[3], p[4]); vop(p[7], p[8]); vop(p[6], p[7]); vop(p[7], p[8]);
2551                     vop(p[10], p[11]); vop(p[9], p[10]); vop(p[10], p[11]); vop(p[6], p[9]); vop(p[8], p[11]);
2552                     vop(p[8], p[9]); vop(p[7], p[10]); vop(p[7], p[8]); vop(p[9], p[10]); vop(p[0], p[6]);
2553                     vop(p[4], p[10]); vop(p[4], p[6]); vop(p[2], p[8]); vop(p[2], p[4]); vop(p[6], p[8]);
2554                     vop(p[1], p[7]); vop(p[5], p[11]); vop(p[5], p[7]); vop(p[3], p[9]); vop(p[3], p[5]);
2555                     vop(p[7], p[9]); vop(p[1], p[2]); vop(p[3], p[4]); vop(p[5], p[6]); vop(p[7], p[8]);
2556                     vop(p[9], p[10]); vop(p[13], p[14]); vop(p[12], p[13]); vop(p[13], p[14]); vop(p[16], p[17]);
2557                     vop(p[15], p[16]); vop(p[16], p[17]); vop(p[12], p[15]); vop(p[14], p[17]); vop(p[14], p[15]);
2558                     vop(p[13], p[16]); vop(p[13], p[14]); vop(p[15], p[16]); vop(p[19], p[20]); vop(p[18], p[19]);
2559                     vop(p[19], p[20]); vop(p[21], p[22]); vop(p[23], p[24]); vop(p[21], p[23]); vop(p[22], p[24]);
2560                     vop(p[22], p[23]); vop(p[18], p[21]); vop(p[20], p[23]); vop(p[20], p[21]); vop(p[19], p[22]);
2561                     vop(p[22], p[24]); vop(p[19], p[20]); vop(p[21], p[22]); vop(p[23], p[24]); vop(p[12], p[18]);
2562                     vop(p[16], p[22]); vop(p[16], p[18]); vop(p[14], p[20]); vop(p[20], p[24]); vop(p[14], p[16]);
2563                     vop(p[18], p[20]); vop(p[22], p[24]); vop(p[13], p[19]); vop(p[17], p[23]); vop(p[17], p[19]);
2564                     vop(p[15], p[21]); vop(p[15], p[17]); vop(p[19], p[21]); vop(p[13], p[14]); vop(p[15], p[16]);
2565                     vop(p[17], p[18]); vop(p[19], p[20]); vop(p[21], p[22]); vop(p[23], p[24]); vop(p[0], p[12]);
2566                     vop(p[8], p[20]); vop(p[8], p[12]); vop(p[4], p[16]); vop(p[16], p[24]); vop(p[12], p[16]);
2567                     vop(p[2], p[14]); vop(p[10], p[22]); vop(p[10], p[14]); vop(p[6], p[18]); vop(p[6], p[10]);
2568                     vop(p[10], p[12]); vop(p[1], p[13]); vop(p[9], p[21]); vop(p[9], p[13]); vop(p[5], p[17]);
2569                     vop(p[13], p[17]); vop(p[3], p[15]); vop(p[11], p[23]); vop(p[11], p[15]); vop(p[7], p[19]);
2570                     vop(p[7], p[11]); vop(p[11], p[13]); vop(p[11], p[12]);
2571                     vop.store(dst+j, p[12]);
2572                 }
2573 
2574                 limit = size.width;
2575             }
2576         }
2577     }
2578 }
2579 
2580 #ifdef HAVE_OPENCL
2581 
ocl_medianFilter(InputArray _src,OutputArray _dst,int m)2582 static bool ocl_medianFilter(InputArray _src, OutputArray _dst, int m)
2583 {
2584     size_t localsize[2] = { 16, 16 };
2585     size_t globalsize[2];
2586     int type = _src.type(), depth = CV_MAT_DEPTH(type), cn = CV_MAT_CN(type);
2587 
2588     if ( !((depth == CV_8U || depth == CV_16U || depth == CV_16S || depth == CV_32F) && cn <= 4 && (m == 3 || m == 5)) )
2589         return false;
2590 
2591     Size imgSize = _src.size();
2592     bool useOptimized = (1 == cn) &&
2593                         (size_t)imgSize.width >= localsize[0] * 8  &&
2594                         (size_t)imgSize.height >= localsize[1] * 8 &&
2595                         imgSize.width % 4 == 0 &&
2596                         imgSize.height % 4 == 0 &&
2597                         (ocl::Device::getDefault().isIntel());
2598 
2599     cv::String kname = format( useOptimized ? "medianFilter%d_u" : "medianFilter%d", m) ;
2600     cv::String kdefs = useOptimized ?
2601                          format("-D T=%s -D T1=%s -D T4=%s%d -D cn=%d -D USE_4OPT", ocl::typeToStr(type),
2602                          ocl::typeToStr(depth), ocl::typeToStr(depth), cn*4, cn)
2603                          :
2604                          format("-D T=%s -D T1=%s -D cn=%d", ocl::typeToStr(type), ocl::typeToStr(depth), cn) ;
2605 
2606     ocl::Kernel k(kname.c_str(), ocl::imgproc::medianFilter_oclsrc, kdefs.c_str() );
2607 
2608     if (k.empty())
2609         return false;
2610 
2611     UMat src = _src.getUMat();
2612     _dst.create(src.size(), type);
2613     UMat dst = _dst.getUMat();
2614 
2615     k.args(ocl::KernelArg::ReadOnlyNoSize(src), ocl::KernelArg::WriteOnly(dst));
2616 
2617     if( useOptimized )
2618     {
2619         globalsize[0] = DIVUP(src.cols / 4, localsize[0]) * localsize[0];
2620         globalsize[1] = DIVUP(src.rows / 4, localsize[1]) * localsize[1];
2621     }
2622     else
2623     {
2624         globalsize[0] = (src.cols + localsize[0] + 2) / localsize[0] * localsize[0];
2625         globalsize[1] = (src.rows + localsize[1] - 1) / localsize[1] * localsize[1];
2626     }
2627 
2628     return k.run(2, globalsize, localsize, false);
2629 }
2630 
2631 #endif
2632 
2633 }
2634 
medianBlur(InputArray _src0,OutputArray _dst,int ksize)2635 void cv::medianBlur( InputArray _src0, OutputArray _dst, int ksize )
2636 {
2637     CV_Assert( (ksize % 2 == 1) && (_src0.dims() <= 2 ));
2638 
2639     if( ksize <= 1 )
2640     {
2641         _src0.copyTo(_dst);
2642         return;
2643     }
2644 
2645     CV_OCL_RUN(_dst.isUMat(),
2646                ocl_medianFilter(_src0,_dst, ksize))
2647 
2648     Mat src0 = _src0.getMat();
2649     _dst.create( src0.size(), src0.type() );
2650     Mat dst = _dst.getMat();
2651 
2652 #if IPP_VERSION_X100 >= 801
2653     CV_IPP_CHECK()
2654     {
2655 #define IPP_FILTER_MEDIAN_BORDER(ippType, ippDataType, flavor) \
2656         do \
2657         { \
2658             if (ippiFilterMedianBorderGetBufferSize(dstRoiSize, maskSize, \
2659                 ippDataType, CV_MAT_CN(type), &bufSize) >= 0) \
2660             { \
2661                 Ipp8u * buffer = ippsMalloc_8u(bufSize); \
2662                 IppStatus status = ippiFilterMedianBorder_##flavor(src.ptr<ippType>(), (int)src.step, \
2663                     dst.ptr<ippType>(), (int)dst.step, dstRoiSize, maskSize, \
2664                     ippBorderRepl, (ippType)0, buffer); \
2665                 ippsFree(buffer); \
2666                 if (status >= 0) \
2667                 { \
2668                     CV_IMPL_ADD(CV_IMPL_IPP); \
2669                     return; \
2670                 } \
2671             } \
2672             setIppErrorStatus(); \
2673         } \
2674         while ((void)0, 0)
2675 
2676         if( ksize <= 5 )
2677         {
2678             Ipp32s bufSize;
2679             IppiSize dstRoiSize = ippiSize(dst.cols, dst.rows), maskSize = ippiSize(ksize, ksize);
2680             Mat src;
2681             if( dst.data != src0.data )
2682                 src = src0;
2683             else
2684                 src0.copyTo(src);
2685 
2686             int type = src0.type();
2687             if (type == CV_8UC1)
2688                 IPP_FILTER_MEDIAN_BORDER(Ipp8u, ipp8u, 8u_C1R);
2689             else if (type == CV_16UC1)
2690                 IPP_FILTER_MEDIAN_BORDER(Ipp16u, ipp16u, 16u_C1R);
2691             else if (type == CV_16SC1)
2692                 IPP_FILTER_MEDIAN_BORDER(Ipp16s, ipp16s, 16s_C1R);
2693             else if (type == CV_32FC1)
2694                 IPP_FILTER_MEDIAN_BORDER(Ipp32f, ipp32f, 32f_C1R);
2695         }
2696 #undef IPP_FILTER_MEDIAN_BORDER
2697     }
2698 #endif
2699 
2700 #ifdef HAVE_TEGRA_OPTIMIZATION
2701     if (tegra::useTegra() && tegra::medianBlur(src0, dst, ksize))
2702         return;
2703 #endif
2704 
2705     bool useSortNet = ksize == 3 || (ksize == 5
2706 #if !(CV_SSE2 || CV_NEON)
2707             && src0.depth() > CV_8U
2708 #endif
2709         );
2710 
2711     Mat src;
2712     if( useSortNet )
2713     {
2714         if( dst.data != src0.data )
2715             src = src0;
2716         else
2717             src0.copyTo(src);
2718 
2719         if( src.depth() == CV_8U )
2720             medianBlur_SortNet<MinMax8u, MinMaxVec8u>( src, dst, ksize );
2721         else if( src.depth() == CV_16U )
2722             medianBlur_SortNet<MinMax16u, MinMaxVec16u>( src, dst, ksize );
2723         else if( src.depth() == CV_16S )
2724             medianBlur_SortNet<MinMax16s, MinMaxVec16s>( src, dst, ksize );
2725         else if( src.depth() == CV_32F )
2726             medianBlur_SortNet<MinMax32f, MinMaxVec32f>( src, dst, ksize );
2727         else
2728             CV_Error(CV_StsUnsupportedFormat, "");
2729 
2730         return;
2731     }
2732     else
2733     {
2734         cv::copyMakeBorder( src0, src, 0, 0, ksize/2, ksize/2, BORDER_REPLICATE );
2735 
2736         int cn = src0.channels();
2737         CV_Assert( src.depth() == CV_8U && (cn == 1 || cn == 3 || cn == 4) );
2738 
2739         double img_size_mp = (double)(src0.total())/(1 << 20);
2740         if( ksize <= 3 + (img_size_mp < 1 ? 12 : img_size_mp < 4 ? 6 : 2)*
2741             (MEDIAN_HAVE_SIMD && (checkHardwareSupport(CV_CPU_SSE2) || checkHardwareSupport(CV_CPU_NEON)) ? 1 : 3))
2742             medianBlur_8u_Om( src, dst, ksize );
2743         else
2744             medianBlur_8u_O1( src, dst, ksize );
2745     }
2746 }
2747 
2748 /****************************************************************************************\
2749                                    Bilateral Filtering
2750 \****************************************************************************************/
2751 
2752 namespace cv
2753 {
2754 
2755 class BilateralFilter_8u_Invoker :
2756     public ParallelLoopBody
2757 {
2758 public:
BilateralFilter_8u_Invoker(Mat & _dest,const Mat & _temp,int _radius,int _maxk,int * _space_ofs,float * _space_weight,float * _color_weight)2759     BilateralFilter_8u_Invoker(Mat& _dest, const Mat& _temp, int _radius, int _maxk,
2760         int* _space_ofs, float *_space_weight, float *_color_weight) :
2761         temp(&_temp), dest(&_dest), radius(_radius),
2762         maxk(_maxk), space_ofs(_space_ofs), space_weight(_space_weight), color_weight(_color_weight)
2763     {
2764     }
2765 
operator ()(const Range & range) const2766     virtual void operator() (const Range& range) const
2767     {
2768         int i, j, cn = dest->channels(), k;
2769         Size size = dest->size();
2770         #if CV_SSE3
2771         int CV_DECL_ALIGNED(16) buf[4];
2772         float CV_DECL_ALIGNED(16) bufSum[4];
2773         static const unsigned int CV_DECL_ALIGNED(16) bufSignMask[] = { 0x80000000, 0x80000000, 0x80000000, 0x80000000 };
2774         bool haveSSE3 = checkHardwareSupport(CV_CPU_SSE3);
2775         #endif
2776 
2777         for( i = range.start; i < range.end; i++ )
2778         {
2779             const uchar* sptr = temp->ptr(i+radius) + radius*cn;
2780             uchar* dptr = dest->ptr(i);
2781 
2782             if( cn == 1 )
2783             {
2784                 for( j = 0; j < size.width; j++ )
2785                 {
2786                     float sum = 0, wsum = 0;
2787                     int val0 = sptr[j];
2788                     k = 0;
2789                     #if CV_SSE3
2790                     if( haveSSE3 )
2791                     {
2792                         __m128 _val0 = _mm_set1_ps(static_cast<float>(val0));
2793                         const __m128 _signMask = _mm_load_ps((const float*)bufSignMask);
2794 
2795                         for( ; k <= maxk - 4; k += 4 )
2796                         {
2797                             __m128 _valF = _mm_set_ps(sptr[j + space_ofs[k+3]], sptr[j + space_ofs[k+2]],
2798                                                       sptr[j + space_ofs[k+1]], sptr[j + space_ofs[k]]);
2799 
2800                             __m128 _val = _mm_andnot_ps(_signMask, _mm_sub_ps(_valF, _val0));
2801                             _mm_store_si128((__m128i*)buf, _mm_cvtps_epi32(_val));
2802 
2803                             __m128 _cw = _mm_set_ps(color_weight[buf[3]],color_weight[buf[2]],
2804                                                     color_weight[buf[1]],color_weight[buf[0]]);
2805                             __m128 _sw = _mm_loadu_ps(space_weight+k);
2806                             __m128 _w = _mm_mul_ps(_cw, _sw);
2807                              _cw = _mm_mul_ps(_w, _valF);
2808 
2809                              _sw = _mm_hadd_ps(_w, _cw);
2810                              _sw = _mm_hadd_ps(_sw, _sw);
2811                              _mm_storel_pi((__m64*)bufSum, _sw);
2812 
2813                              sum += bufSum[1];
2814                              wsum += bufSum[0];
2815                         }
2816                     }
2817                     #endif
2818                     for( ; k < maxk; k++ )
2819                     {
2820                         int val = sptr[j + space_ofs[k]];
2821                         float w = space_weight[k]*color_weight[std::abs(val - val0)];
2822                         sum += val*w;
2823                         wsum += w;
2824                     }
2825                     // overflow is not possible here => there is no need to use cv::saturate_cast
2826                     dptr[j] = (uchar)cvRound(sum/wsum);
2827                 }
2828             }
2829             else
2830             {
2831                 assert( cn == 3 );
2832                 for( j = 0; j < size.width*3; j += 3 )
2833                 {
2834                     float sum_b = 0, sum_g = 0, sum_r = 0, wsum = 0;
2835                     int b0 = sptr[j], g0 = sptr[j+1], r0 = sptr[j+2];
2836                     k = 0;
2837                     #if CV_SSE3
2838                     if( haveSSE3 )
2839                     {
2840                         const __m128i izero = _mm_setzero_si128();
2841                         const __m128 _b0 = _mm_set1_ps(static_cast<float>(b0));
2842                         const __m128 _g0 = _mm_set1_ps(static_cast<float>(g0));
2843                         const __m128 _r0 = _mm_set1_ps(static_cast<float>(r0));
2844                         const __m128 _signMask = _mm_load_ps((const float*)bufSignMask);
2845 
2846                         for( ; k <= maxk - 4; k += 4 )
2847                         {
2848                             const int* const sptr_k0  = reinterpret_cast<const int*>(sptr + j + space_ofs[k]);
2849                             const int* const sptr_k1  = reinterpret_cast<const int*>(sptr + j + space_ofs[k+1]);
2850                             const int* const sptr_k2  = reinterpret_cast<const int*>(sptr + j + space_ofs[k+2]);
2851                             const int* const sptr_k3  = reinterpret_cast<const int*>(sptr + j + space_ofs[k+3]);
2852 
2853                             __m128 _b = _mm_cvtepi32_ps(_mm_unpacklo_epi16(_mm_unpacklo_epi8(_mm_cvtsi32_si128(sptr_k0[0]), izero), izero));
2854                             __m128 _g = _mm_cvtepi32_ps(_mm_unpacklo_epi16(_mm_unpacklo_epi8(_mm_cvtsi32_si128(sptr_k1[0]), izero), izero));
2855                             __m128 _r = _mm_cvtepi32_ps(_mm_unpacklo_epi16(_mm_unpacklo_epi8(_mm_cvtsi32_si128(sptr_k2[0]), izero), izero));
2856                             __m128 _z = _mm_cvtepi32_ps(_mm_unpacklo_epi16(_mm_unpacklo_epi8(_mm_cvtsi32_si128(sptr_k3[0]), izero), izero));
2857 
2858                             _MM_TRANSPOSE4_PS(_b, _g, _r, _z);
2859 
2860                             __m128 bt = _mm_andnot_ps(_signMask, _mm_sub_ps(_b,_b0));
2861                             __m128 gt = _mm_andnot_ps(_signMask, _mm_sub_ps(_g,_g0));
2862                             __m128 rt = _mm_andnot_ps(_signMask, _mm_sub_ps(_r,_r0));
2863 
2864                             bt =_mm_add_ps(rt, _mm_add_ps(bt, gt));
2865                             _mm_store_si128((__m128i*)buf, _mm_cvtps_epi32(bt));
2866 
2867                             __m128 _w  = _mm_set_ps(color_weight[buf[3]],color_weight[buf[2]],
2868                                                     color_weight[buf[1]],color_weight[buf[0]]);
2869                             __m128 _sw = _mm_loadu_ps(space_weight+k);
2870 
2871                             _w = _mm_mul_ps(_w,_sw);
2872                             _b = _mm_mul_ps(_b, _w);
2873                             _g = _mm_mul_ps(_g, _w);
2874                             _r = _mm_mul_ps(_r, _w);
2875 
2876                              _w = _mm_hadd_ps(_w, _b);
2877                              _g = _mm_hadd_ps(_g, _r);
2878 
2879                              _w = _mm_hadd_ps(_w, _g);
2880                              _mm_store_ps(bufSum, _w);
2881 
2882                              wsum  += bufSum[0];
2883                              sum_b += bufSum[1];
2884                              sum_g += bufSum[2];
2885                              sum_r += bufSum[3];
2886                          }
2887                     }
2888                     #endif
2889 
2890                     for( ; k < maxk; k++ )
2891                     {
2892                         const uchar* sptr_k = sptr + j + space_ofs[k];
2893                         int b = sptr_k[0], g = sptr_k[1], r = sptr_k[2];
2894                         float w = space_weight[k]*color_weight[std::abs(b - b0) +
2895                                                                std::abs(g - g0) + std::abs(r - r0)];
2896                         sum_b += b*w; sum_g += g*w; sum_r += r*w;
2897                         wsum += w;
2898                     }
2899                     wsum = 1.f/wsum;
2900                     b0 = cvRound(sum_b*wsum);
2901                     g0 = cvRound(sum_g*wsum);
2902                     r0 = cvRound(sum_r*wsum);
2903                     dptr[j] = (uchar)b0; dptr[j+1] = (uchar)g0; dptr[j+2] = (uchar)r0;
2904                 }
2905             }
2906         }
2907     }
2908 
2909 private:
2910     const Mat *temp;
2911     Mat *dest;
2912     int radius, maxk, *space_ofs;
2913     float *space_weight, *color_weight;
2914 };
2915 
2916 #if defined (HAVE_IPP) && !defined(HAVE_IPP_ICV_ONLY) && 0
2917 class IPPBilateralFilter_8u_Invoker :
2918     public ParallelLoopBody
2919 {
2920 public:
IPPBilateralFilter_8u_Invoker(Mat & _src,Mat & _dst,double _sigma_color,double _sigma_space,int _radius,bool * _ok)2921     IPPBilateralFilter_8u_Invoker(Mat &_src, Mat &_dst, double _sigma_color, double _sigma_space, int _radius, bool *_ok) :
2922       ParallelLoopBody(), src(_src), dst(_dst), sigma_color(_sigma_color), sigma_space(_sigma_space), radius(_radius), ok(_ok)
2923       {
2924           *ok = true;
2925       }
2926 
operator ()(const Range & range) const2927       virtual void operator() (const Range& range) const
2928       {
2929           int d = radius * 2 + 1;
2930           IppiSize kernel = {d, d};
2931           IppiSize roi={dst.cols, range.end - range.start};
2932           int bufsize=0;
2933           if (0 > ippiFilterBilateralGetBufSize_8u_C1R( ippiFilterBilateralGauss, roi, kernel, &bufsize))
2934           {
2935               *ok = false;
2936               return;
2937           }
2938           AutoBuffer<uchar> buf(bufsize);
2939           IppiFilterBilateralSpec *pSpec = (IppiFilterBilateralSpec *)alignPtr(&buf[0], 32);
2940           if (0 > ippiFilterBilateralInit_8u_C1R( ippiFilterBilateralGauss, kernel, (Ipp32f)sigma_color, (Ipp32f)sigma_space, 1, pSpec ))
2941           {
2942               *ok = false;
2943               return;
2944           }
2945           if (0 > ippiFilterBilateral_8u_C1R( src.ptr<uchar>(range.start) + radius * ((int)src.step[0] + 1), (int)src.step[0], dst.ptr<uchar>(range.start), (int)dst.step[0], roi, kernel, pSpec ))
2946               *ok = false;
2947           else
2948           {
2949             CV_IMPL_ADD(CV_IMPL_IPP|CV_IMPL_MT);
2950           }
2951       }
2952 private:
2953     Mat &src;
2954     Mat &dst;
2955     double sigma_color;
2956     double sigma_space;
2957     int radius;
2958     bool *ok;
2959     const IPPBilateralFilter_8u_Invoker& operator= (const IPPBilateralFilter_8u_Invoker&);
2960 };
2961 #endif
2962 
2963 #ifdef HAVE_OPENCL
2964 
ocl_bilateralFilter_8u(InputArray _src,OutputArray _dst,int d,double sigma_color,double sigma_space,int borderType)2965 static bool ocl_bilateralFilter_8u(InputArray _src, OutputArray _dst, int d,
2966                                    double sigma_color, double sigma_space,
2967                                    int borderType)
2968 {
2969 #ifdef ANDROID
2970     if (ocl::Device::getDefault().isNVidia())
2971         return false;
2972 #endif
2973 
2974     int type = _src.type(), depth = CV_MAT_DEPTH(type), cn = CV_MAT_CN(type);
2975     int i, j, maxk, radius;
2976 
2977     if (depth != CV_8U || cn > 4)
2978         return false;
2979 
2980     if (sigma_color <= 0)
2981         sigma_color = 1;
2982     if (sigma_space <= 0)
2983         sigma_space = 1;
2984 
2985     double gauss_color_coeff = -0.5 / (sigma_color * sigma_color);
2986     double gauss_space_coeff = -0.5 / (sigma_space * sigma_space);
2987 
2988     if ( d <= 0 )
2989         radius = cvRound(sigma_space * 1.5);
2990     else
2991         radius = d / 2;
2992     radius = MAX(radius, 1);
2993     d = radius * 2 + 1;
2994 
2995     UMat src = _src.getUMat(), dst = _dst.getUMat(), temp;
2996     if (src.u == dst.u)
2997         return false;
2998 
2999     copyMakeBorder(src, temp, radius, radius, radius, radius, borderType);
3000     std::vector<float> _space_weight(d * d);
3001     std::vector<int> _space_ofs(d * d);
3002     float * const space_weight = &_space_weight[0];
3003     int * const space_ofs = &_space_ofs[0];
3004 
3005     // initialize space-related bilateral filter coefficients
3006     for( i = -radius, maxk = 0; i <= radius; i++ )
3007         for( j = -radius; j <= radius; j++ )
3008         {
3009             double r = std::sqrt((double)i * i + (double)j * j);
3010             if ( r > radius )
3011                 continue;
3012             space_weight[maxk] = (float)std::exp(r * r * gauss_space_coeff);
3013             space_ofs[maxk++] = (int)(i * temp.step + j * cn);
3014         }
3015 
3016     char cvt[3][40];
3017     String cnstr = cn > 1 ? format("%d", cn) : "";
3018     String kernelName("bilateral");
3019     size_t sizeDiv = 1;
3020     if ((ocl::Device::getDefault().isIntel()) &&
3021         (ocl::Device::getDefault().type() == ocl::Device::TYPE_GPU))
3022     {
3023             //Intel GPU
3024             if (dst.cols % 4 == 0 && cn == 1) // For single channel x4 sized images.
3025             {
3026                 kernelName = "bilateral_float4";
3027                 sizeDiv = 4;
3028             }
3029      }
3030      ocl::Kernel k(kernelName.c_str(), ocl::imgproc::bilateral_oclsrc,
3031             format("-D radius=%d -D maxk=%d -D cn=%d -D int_t=%s -D uint_t=uint%s -D convert_int_t=%s"
3032             " -D uchar_t=%s -D float_t=%s -D convert_float_t=%s -D convert_uchar_t=%s -D gauss_color_coeff=%f",
3033             radius, maxk, cn, ocl::typeToStr(CV_32SC(cn)), cnstr.c_str(),
3034             ocl::convertTypeStr(CV_8U, CV_32S, cn, cvt[0]),
3035             ocl::typeToStr(type), ocl::typeToStr(CV_32FC(cn)),
3036             ocl::convertTypeStr(CV_32S, CV_32F, cn, cvt[1]),
3037             ocl::convertTypeStr(CV_32F, CV_8U, cn, cvt[2]), gauss_color_coeff));
3038     if (k.empty())
3039         return false;
3040 
3041     Mat mspace_weight(1, d * d, CV_32FC1, space_weight);
3042     Mat mspace_ofs(1, d * d, CV_32SC1, space_ofs);
3043     UMat ucolor_weight, uspace_weight, uspace_ofs;
3044 
3045     mspace_weight.copyTo(uspace_weight);
3046     mspace_ofs.copyTo(uspace_ofs);
3047 
3048     k.args(ocl::KernelArg::ReadOnlyNoSize(temp), ocl::KernelArg::WriteOnly(dst),
3049            ocl::KernelArg::PtrReadOnly(uspace_weight),
3050            ocl::KernelArg::PtrReadOnly(uspace_ofs));
3051 
3052     size_t globalsize[2] = { dst.cols / sizeDiv, dst.rows };
3053     return k.run(2, globalsize, NULL, false);
3054 }
3055 
3056 #endif
3057 static void
bilateralFilter_8u(const Mat & src,Mat & dst,int d,double sigma_color,double sigma_space,int borderType)3058 bilateralFilter_8u( const Mat& src, Mat& dst, int d,
3059     double sigma_color, double sigma_space,
3060     int borderType )
3061 {
3062     int cn = src.channels();
3063     int i, j, maxk, radius;
3064     Size size = src.size();
3065 
3066     CV_Assert( (src.type() == CV_8UC1 || src.type() == CV_8UC3) && src.data != dst.data );
3067 
3068     if( sigma_color <= 0 )
3069         sigma_color = 1;
3070     if( sigma_space <= 0 )
3071         sigma_space = 1;
3072 
3073     double gauss_color_coeff = -0.5/(sigma_color*sigma_color);
3074     double gauss_space_coeff = -0.5/(sigma_space*sigma_space);
3075 
3076     if( d <= 0 )
3077         radius = cvRound(sigma_space*1.5);
3078     else
3079         radius = d/2;
3080     radius = MAX(radius, 1);
3081     d = radius*2 + 1;
3082 
3083     Mat temp;
3084     copyMakeBorder( src, temp, radius, radius, radius, radius, borderType );
3085 
3086 #if defined HAVE_IPP && (IPP_VERSION_MAJOR >= 7) && 0
3087     CV_IPP_CHECK()
3088     {
3089         if( cn == 1 )
3090         {
3091             bool ok;
3092             IPPBilateralFilter_8u_Invoker body(temp, dst, sigma_color * sigma_color, sigma_space * sigma_space, radius, &ok );
3093             parallel_for_(Range(0, dst.rows), body, dst.total()/(double)(1<<16));
3094             if( ok )
3095             {
3096                 CV_IMPL_ADD(CV_IMPL_IPP|CV_IMPL_MT);
3097                 return;
3098             }
3099             setIppErrorStatus();
3100         }
3101     }
3102 #endif
3103 
3104     std::vector<float> _color_weight(cn*256);
3105     std::vector<float> _space_weight(d*d);
3106     std::vector<int> _space_ofs(d*d);
3107     float* color_weight = &_color_weight[0];
3108     float* space_weight = &_space_weight[0];
3109     int* space_ofs = &_space_ofs[0];
3110 
3111     // initialize color-related bilateral filter coefficients
3112 
3113     for( i = 0; i < 256*cn; i++ )
3114         color_weight[i] = (float)std::exp(i*i*gauss_color_coeff);
3115 
3116     // initialize space-related bilateral filter coefficients
3117     for( i = -radius, maxk = 0; i <= radius; i++ )
3118     {
3119         j = -radius;
3120 
3121         for( ; j <= radius; j++ )
3122         {
3123             double r = std::sqrt((double)i*i + (double)j*j);
3124             if( r > radius )
3125                 continue;
3126             space_weight[maxk] = (float)std::exp(r*r*gauss_space_coeff);
3127             space_ofs[maxk++] = (int)(i*temp.step + j*cn);
3128         }
3129     }
3130 
3131     BilateralFilter_8u_Invoker body(dst, temp, radius, maxk, space_ofs, space_weight, color_weight);
3132     parallel_for_(Range(0, size.height), body, dst.total()/(double)(1<<16));
3133 }
3134 
3135 
3136 class BilateralFilter_32f_Invoker :
3137     public ParallelLoopBody
3138 {
3139 public:
3140 
BilateralFilter_32f_Invoker(int _cn,int _radius,int _maxk,int * _space_ofs,const Mat & _temp,Mat & _dest,float _scale_index,float * _space_weight,float * _expLUT)3141     BilateralFilter_32f_Invoker(int _cn, int _radius, int _maxk, int *_space_ofs,
3142         const Mat& _temp, Mat& _dest, float _scale_index, float *_space_weight, float *_expLUT) :
3143         cn(_cn), radius(_radius), maxk(_maxk), space_ofs(_space_ofs),
3144         temp(&_temp), dest(&_dest), scale_index(_scale_index), space_weight(_space_weight), expLUT(_expLUT)
3145     {
3146     }
3147 
operator ()(const Range & range) const3148     virtual void operator() (const Range& range) const
3149     {
3150         int i, j, k;
3151         Size size = dest->size();
3152         #if CV_SSE3
3153         int CV_DECL_ALIGNED(16) idxBuf[4];
3154         float CV_DECL_ALIGNED(16) bufSum32[4];
3155         static const unsigned int CV_DECL_ALIGNED(16) bufSignMask[] = { 0x80000000, 0x80000000, 0x80000000, 0x80000000 };
3156         bool haveSSE3 = checkHardwareSupport(CV_CPU_SSE3);
3157         #endif
3158 
3159         for( i = range.start; i < range.end; i++ )
3160         {
3161             const float* sptr = temp->ptr<float>(i+radius) + radius*cn;
3162             float* dptr = dest->ptr<float>(i);
3163 
3164             if( cn == 1 )
3165             {
3166                 for( j = 0; j < size.width; j++ )
3167                 {
3168                     float sum = 0, wsum = 0;
3169                     float val0 = sptr[j];
3170                     k = 0;
3171                     #if CV_SSE3
3172                     if( haveSSE3 )
3173                     {
3174                         __m128 psum = _mm_setzero_ps();
3175                         const __m128 _val0 = _mm_set1_ps(sptr[j]);
3176                         const __m128 _scale_index = _mm_set1_ps(scale_index);
3177                         const __m128 _signMask = _mm_load_ps((const float*)bufSignMask);
3178 
3179                         for( ; k <= maxk - 4 ; k += 4 )
3180                         {
3181                             __m128 _sw    = _mm_loadu_ps(space_weight + k);
3182                             __m128 _val   = _mm_set_ps(sptr[j + space_ofs[k+3]], sptr[j + space_ofs[k+2]],
3183                                                        sptr[j + space_ofs[k+1]], sptr[j + space_ofs[k]]);
3184                             __m128 _alpha = _mm_mul_ps(_mm_andnot_ps( _signMask, _mm_sub_ps(_val,_val0)), _scale_index);
3185 
3186                             __m128i _idx = _mm_cvtps_epi32(_alpha);
3187                             _mm_store_si128((__m128i*)idxBuf, _idx);
3188                             _alpha = _mm_sub_ps(_alpha, _mm_cvtepi32_ps(_idx));
3189 
3190                             __m128 _explut  = _mm_set_ps(expLUT[idxBuf[3]], expLUT[idxBuf[2]],
3191                                                          expLUT[idxBuf[1]], expLUT[idxBuf[0]]);
3192                             __m128 _explut1 = _mm_set_ps(expLUT[idxBuf[3]+1], expLUT[idxBuf[2]+1],
3193                                                          expLUT[idxBuf[1]+1], expLUT[idxBuf[0]+1]);
3194 
3195                             __m128 _w = _mm_mul_ps(_sw, _mm_add_ps(_explut, _mm_mul_ps(_alpha, _mm_sub_ps(_explut1, _explut))));
3196                             _val = _mm_mul_ps(_w, _val);
3197 
3198                              _sw = _mm_hadd_ps(_w, _val);
3199                              _sw = _mm_hadd_ps(_sw, _sw);
3200                              psum = _mm_add_ps(_sw, psum);
3201                         }
3202                         _mm_storel_pi((__m64*)bufSum32, psum);
3203 
3204                         sum = bufSum32[1];
3205                         wsum = bufSum32[0];
3206                     }
3207                     #endif
3208 
3209                     for( ; k < maxk; k++ )
3210                     {
3211                         float val = sptr[j + space_ofs[k]];
3212                         float alpha = (float)(std::abs(val - val0)*scale_index);
3213                         int idx = cvFloor(alpha);
3214                         alpha -= idx;
3215                         float w = space_weight[k]*(expLUT[idx] + alpha*(expLUT[idx+1] - expLUT[idx]));
3216                         sum += val*w;
3217                         wsum += w;
3218                     }
3219                     dptr[j] = (float)(sum/wsum);
3220                 }
3221             }
3222             else
3223             {
3224                 CV_Assert( cn == 3 );
3225                 for( j = 0; j < size.width*3; j += 3 )
3226                 {
3227                     float sum_b = 0, sum_g = 0, sum_r = 0, wsum = 0;
3228                     float b0 = sptr[j], g0 = sptr[j+1], r0 = sptr[j+2];
3229                     k = 0;
3230                     #if  CV_SSE3
3231                     if( haveSSE3 )
3232                     {
3233                         __m128 sum = _mm_setzero_ps();
3234                         const __m128 _b0 = _mm_set1_ps(b0);
3235                         const __m128 _g0 = _mm_set1_ps(g0);
3236                         const __m128 _r0 = _mm_set1_ps(r0);
3237                         const __m128 _scale_index = _mm_set1_ps(scale_index);
3238                         const __m128 _signMask = _mm_load_ps((const float*)bufSignMask);
3239 
3240                         for( ; k <= maxk-4; k += 4 )
3241                         {
3242                             __m128 _sw = _mm_loadu_ps(space_weight + k);
3243 
3244                             const float* const sptr_k0 = sptr + j + space_ofs[k];
3245                             const float* const sptr_k1 = sptr + j + space_ofs[k+1];
3246                             const float* const sptr_k2 = sptr + j + space_ofs[k+2];
3247                             const float* const sptr_k3 = sptr + j + space_ofs[k+3];
3248 
3249                             __m128 _b = _mm_loadu_ps(sptr_k0);
3250                             __m128 _g = _mm_loadu_ps(sptr_k1);
3251                             __m128 _r = _mm_loadu_ps(sptr_k2);
3252                             __m128 _z = _mm_loadu_ps(sptr_k3);
3253                             _MM_TRANSPOSE4_PS(_b, _g, _r, _z);
3254 
3255                             __m128 _bt = _mm_andnot_ps(_signMask,_mm_sub_ps(_b,_b0));
3256                             __m128 _gt = _mm_andnot_ps(_signMask,_mm_sub_ps(_g,_g0));
3257                             __m128 _rt = _mm_andnot_ps(_signMask,_mm_sub_ps(_r,_r0));
3258 
3259                             __m128 _alpha = _mm_mul_ps(_scale_index, _mm_add_ps(_rt,_mm_add_ps(_bt, _gt)));
3260 
3261                             __m128i _idx  = _mm_cvtps_epi32(_alpha);
3262                             _mm_store_si128((__m128i*)idxBuf, _idx);
3263                             _alpha = _mm_sub_ps(_alpha, _mm_cvtepi32_ps(_idx));
3264 
3265                             __m128 _explut  = _mm_set_ps(expLUT[idxBuf[3]], expLUT[idxBuf[2]], expLUT[idxBuf[1]], expLUT[idxBuf[0]]);
3266                             __m128 _explut1 = _mm_set_ps(expLUT[idxBuf[3]+1], expLUT[idxBuf[2]+1], expLUT[idxBuf[1]+1], expLUT[idxBuf[0]+1]);
3267 
3268                             __m128 _w = _mm_mul_ps(_sw, _mm_add_ps(_explut, _mm_mul_ps(_alpha, _mm_sub_ps(_explut1, _explut))));
3269 
3270                             _b = _mm_mul_ps(_b, _w);
3271                             _g = _mm_mul_ps(_g, _w);
3272                             _r = _mm_mul_ps(_r, _w);
3273 
3274                              _w = _mm_hadd_ps(_w, _b);
3275                              _g = _mm_hadd_ps(_g, _r);
3276 
3277                              _w = _mm_hadd_ps(_w, _g);
3278                              sum = _mm_add_ps(sum, _w);
3279                         }
3280                         _mm_store_ps(bufSum32, sum);
3281                         wsum  = bufSum32[0];
3282                         sum_b = bufSum32[1];
3283                         sum_g = bufSum32[2];
3284                         sum_r = bufSum32[3];
3285                     }
3286                     #endif
3287 
3288                     for(; k < maxk; k++ )
3289                     {
3290                         const float* sptr_k = sptr + j + space_ofs[k];
3291                         float b = sptr_k[0], g = sptr_k[1], r = sptr_k[2];
3292                         float alpha = (float)((std::abs(b - b0) +
3293                             std::abs(g - g0) + std::abs(r - r0))*scale_index);
3294                         int idx = cvFloor(alpha);
3295                         alpha -= idx;
3296                         float w = space_weight[k]*(expLUT[idx] + alpha*(expLUT[idx+1] - expLUT[idx]));
3297                         sum_b += b*w; sum_g += g*w; sum_r += r*w;
3298                         wsum += w;
3299                     }
3300                     wsum = 1.f/wsum;
3301                     b0 = sum_b*wsum;
3302                     g0 = sum_g*wsum;
3303                     r0 = sum_r*wsum;
3304                     dptr[j] = b0; dptr[j+1] = g0; dptr[j+2] = r0;
3305                 }
3306             }
3307         }
3308     }
3309 
3310 private:
3311     int cn, radius, maxk, *space_ofs;
3312     const Mat* temp;
3313     Mat *dest;
3314     float scale_index, *space_weight, *expLUT;
3315 };
3316 
3317 
3318 static void
bilateralFilter_32f(const Mat & src,Mat & dst,int d,double sigma_color,double sigma_space,int borderType)3319 bilateralFilter_32f( const Mat& src, Mat& dst, int d,
3320                      double sigma_color, double sigma_space,
3321                      int borderType )
3322 {
3323     int cn = src.channels();
3324     int i, j, maxk, radius;
3325     double minValSrc=-1, maxValSrc=1;
3326     const int kExpNumBinsPerChannel = 1 << 12;
3327     int kExpNumBins = 0;
3328     float lastExpVal = 1.f;
3329     float len, scale_index;
3330     Size size = src.size();
3331 
3332     CV_Assert( (src.type() == CV_32FC1 || src.type() == CV_32FC3) && src.data != dst.data );
3333 
3334     if( sigma_color <= 0 )
3335         sigma_color = 1;
3336     if( sigma_space <= 0 )
3337         sigma_space = 1;
3338 
3339     double gauss_color_coeff = -0.5/(sigma_color*sigma_color);
3340     double gauss_space_coeff = -0.5/(sigma_space*sigma_space);
3341 
3342     if( d <= 0 )
3343         radius = cvRound(sigma_space*1.5);
3344     else
3345         radius = d/2;
3346     radius = MAX(radius, 1);
3347     d = radius*2 + 1;
3348     // compute the min/max range for the input image (even if multichannel)
3349 
3350     minMaxLoc( src.reshape(1), &minValSrc, &maxValSrc );
3351     if(std::abs(minValSrc - maxValSrc) < FLT_EPSILON)
3352     {
3353         src.copyTo(dst);
3354         return;
3355     }
3356 
3357     // temporary copy of the image with borders for easy processing
3358     Mat temp;
3359     copyMakeBorder( src, temp, radius, radius, radius, radius, borderType );
3360     const double insteadNaNValue = -5. * sigma_color;
3361     patchNaNs( temp, insteadNaNValue ); // this replacement of NaNs makes the assumption that depth values are nonnegative
3362                                         // TODO: make insteadNaNValue avalible in the outside function interface to control the cases breaking the assumption
3363     // allocate lookup tables
3364     std::vector<float> _space_weight(d*d);
3365     std::vector<int> _space_ofs(d*d);
3366     float* space_weight = &_space_weight[0];
3367     int* space_ofs = &_space_ofs[0];
3368 
3369     // assign a length which is slightly more than needed
3370     len = (float)(maxValSrc - minValSrc) * cn;
3371     kExpNumBins = kExpNumBinsPerChannel * cn;
3372     std::vector<float> _expLUT(kExpNumBins+2);
3373     float* expLUT = &_expLUT[0];
3374 
3375     scale_index = kExpNumBins/len;
3376 
3377     // initialize the exp LUT
3378     for( i = 0; i < kExpNumBins+2; i++ )
3379     {
3380         if( lastExpVal > 0.f )
3381         {
3382             double val =  i / scale_index;
3383             expLUT[i] = (float)std::exp(val * val * gauss_color_coeff);
3384             lastExpVal = expLUT[i];
3385         }
3386         else
3387             expLUT[i] = 0.f;
3388     }
3389 
3390     // initialize space-related bilateral filter coefficients
3391     for( i = -radius, maxk = 0; i <= radius; i++ )
3392         for( j = -radius; j <= radius; j++ )
3393         {
3394             double r = std::sqrt((double)i*i + (double)j*j);
3395             if( r > radius )
3396                 continue;
3397             space_weight[maxk] = (float)std::exp(r*r*gauss_space_coeff);
3398             space_ofs[maxk++] = (int)(i*(temp.step/sizeof(float)) + j*cn);
3399         }
3400 
3401     // parallel_for usage
3402 
3403     BilateralFilter_32f_Invoker body(cn, radius, maxk, space_ofs, temp, dst, scale_index, space_weight, expLUT);
3404     parallel_for_(Range(0, size.height), body, dst.total()/(double)(1<<16));
3405 }
3406 
3407 }
3408 
bilateralFilter(InputArray _src,OutputArray _dst,int d,double sigmaColor,double sigmaSpace,int borderType)3409 void cv::bilateralFilter( InputArray _src, OutputArray _dst, int d,
3410                       double sigmaColor, double sigmaSpace,
3411                       int borderType )
3412 {
3413     _dst.create( _src.size(), _src.type() );
3414 
3415     CV_OCL_RUN(_src.dims() <= 2 && _dst.isUMat(),
3416                ocl_bilateralFilter_8u(_src, _dst, d, sigmaColor, sigmaSpace, borderType))
3417 
3418     Mat src = _src.getMat(), dst = _dst.getMat();
3419 
3420     if( src.depth() == CV_8U )
3421         bilateralFilter_8u( src, dst, d, sigmaColor, sigmaSpace, borderType );
3422     else if( src.depth() == CV_32F )
3423         bilateralFilter_32f( src, dst, d, sigmaColor, sigmaSpace, borderType );
3424     else
3425         CV_Error( CV_StsUnsupportedFormat,
3426         "Bilateral filtering is only implemented for 8u and 32f images" );
3427 }
3428 
3429 //////////////////////////////////////////////////////////////////////////////////////////
3430 
3431 CV_IMPL void
cvSmooth(const void * srcarr,void * dstarr,int smooth_type,int param1,int param2,double param3,double param4)3432 cvSmooth( const void* srcarr, void* dstarr, int smooth_type,
3433           int param1, int param2, double param3, double param4 )
3434 {
3435     cv::Mat src = cv::cvarrToMat(srcarr), dst0 = cv::cvarrToMat(dstarr), dst = dst0;
3436 
3437     CV_Assert( dst.size() == src.size() &&
3438         (smooth_type == CV_BLUR_NO_SCALE || dst.type() == src.type()) );
3439 
3440     if( param2 <= 0 )
3441         param2 = param1;
3442 
3443     if( smooth_type == CV_BLUR || smooth_type == CV_BLUR_NO_SCALE )
3444         cv::boxFilter( src, dst, dst.depth(), cv::Size(param1, param2), cv::Point(-1,-1),
3445             smooth_type == CV_BLUR, cv::BORDER_REPLICATE );
3446     else if( smooth_type == CV_GAUSSIAN )
3447         cv::GaussianBlur( src, dst, cv::Size(param1, param2), param3, param4, cv::BORDER_REPLICATE );
3448     else if( smooth_type == CV_MEDIAN )
3449         cv::medianBlur( src, dst, param1 );
3450     else
3451         cv::bilateralFilter( src, dst, param1, param3, param4, cv::BORDER_REPLICATE );
3452 
3453     if( dst.data != dst0.data )
3454         CV_Error( CV_StsUnmatchedFormats, "The destination image does not have the proper type" );
3455 }
3456 
3457 /* End of file. */
3458