• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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, Intel Corporation, all rights reserved.
14 // Copyright (C) 2013, OpenCV Foundation, all rights reserved.
15 // Third party copyrights are property of their respective owners.
16 //
17 // Redistribution and use in source and binary forms, with or without modification,
18 // are permitted provided that the following conditions are met:
19 //
20 //   * Redistribution's of source code must retain the above copyright notice,
21 //     this list of conditions and the following disclaimer.
22 //
23 //   * Redistribution's in binary form must reproduce the above copyright notice,
24 //     this list of conditions and the following disclaimer in the documentation
25 //     and/or other materials provided with the distribution.
26 //
27 //   * The name of the copyright holders may not be used to endorse or promote products
28 //     derived from this software without specific prior written permission.
29 //
30 // This software is provided by the copyright holders and contributors "as is" and
31 // any express or implied warranties, including, but not limited to, the implied
32 // warranties of merchantability and fitness for a particular purpose are disclaimed.
33 // In no event shall the Intel Corporation or contributors be liable for any direct,
34 // indirect, incidental, special, exemplary, or consequential damages
35 // (including, but not limited to, procurement of substitute goods or services;
36 // loss of use, data, or profits; or business interruption) however caused
37 // and on any theory of liability, whether in contract, strict liability,
38 // or tort (including negligence or otherwise) arising in any way out of
39 // the use of this software, even if advised of the possibility of such damage.
40 //
41 //M*/
42 
43 /****************************************************************************************\
44 *    Very fast SAD-based (Sum-of-Absolute-Diffrences) stereo correspondence algorithm.   *
45 *    Contributed by Kurt Konolige                                                        *
46 \****************************************************************************************/
47 
48 #include "precomp.hpp"
49 #include <stdio.h>
50 #include <limits>
51 #include "opencl_kernels_calib3d.hpp"
52 
53 namespace cv
54 {
55 
56 struct StereoBMParams
57 {
StereoBMParamscv::StereoBMParams58     StereoBMParams(int _numDisparities=64, int _SADWindowSize=21)
59     {
60         preFilterType = StereoBM::PREFILTER_XSOBEL;
61         preFilterSize = 9;
62         preFilterCap = 31;
63         SADWindowSize = _SADWindowSize;
64         minDisparity = 0;
65         numDisparities = _numDisparities > 0 ? _numDisparities : 64;
66         textureThreshold = 10;
67         uniquenessRatio = 15;
68         speckleRange = speckleWindowSize = 0;
69         roi1 = roi2 = Rect(0,0,0,0);
70         disp12MaxDiff = -1;
71         dispType = CV_16S;
72     }
73 
74     int preFilterType;
75     int preFilterSize;
76     int preFilterCap;
77     int SADWindowSize;
78     int minDisparity;
79     int numDisparities;
80     int textureThreshold;
81     int uniquenessRatio;
82     int speckleRange;
83     int speckleWindowSize;
84     Rect roi1, roi2;
85     int disp12MaxDiff;
86     int dispType;
87 };
88 
ocl_prefilter_norm(InputArray _input,OutputArray _output,int winsize,int prefilterCap)89 static bool ocl_prefilter_norm(InputArray _input, OutputArray _output, int winsize, int prefilterCap)
90 {
91     ocl::Kernel k("prefilter_norm", ocl::calib3d::stereobm_oclsrc, cv::format("-D WSZ=%d", winsize));
92     if(k.empty())
93         return false;
94 
95     int scale_g = winsize*winsize/8, scale_s = (1024 + scale_g)/(scale_g*2);
96     scale_g *= scale_s;
97 
98     UMat input = _input.getUMat(), output;
99     _output.create(input.size(), input.type());
100     output = _output.getUMat();
101 
102     size_t globalThreads[3] = { input.cols, input.rows, 1 };
103 
104     k.args(ocl::KernelArg::PtrReadOnly(input), ocl::KernelArg::PtrWriteOnly(output), input.rows, input.cols,
105         prefilterCap, scale_g, scale_s);
106 
107     return k.run(2, globalThreads, NULL, false);
108 }
109 
prefilterNorm(const Mat & src,Mat & dst,int winsize,int ftzero,uchar * buf)110 static void prefilterNorm( const Mat& src, Mat& dst, int winsize, int ftzero, uchar* buf )
111 {
112     int x, y, wsz2 = winsize/2;
113     int* vsum = (int*)alignPtr(buf + (wsz2 + 1)*sizeof(vsum[0]), 32);
114     int scale_g = winsize*winsize/8, scale_s = (1024 + scale_g)/(scale_g*2);
115     const int OFS = 256*5, TABSZ = OFS*2 + 256;
116     uchar tab[TABSZ];
117     const uchar* sptr = src.ptr();
118     int srcstep = (int)src.step;
119     Size size = src.size();
120 
121     scale_g *= scale_s;
122 
123     for( x = 0; x < TABSZ; x++ )
124         tab[x] = (uchar)(x - OFS < -ftzero ? 0 : x - OFS > ftzero ? ftzero*2 : x - OFS + ftzero);
125 
126     for( x = 0; x < size.width; x++ )
127         vsum[x] = (ushort)(sptr[x]*(wsz2 + 2));
128 
129     for( y = 1; y < wsz2; y++ )
130     {
131         for( x = 0; x < size.width; x++ )
132             vsum[x] = (ushort)(vsum[x] + sptr[srcstep*y + x]);
133     }
134 
135     for( y = 0; y < size.height; y++ )
136     {
137         const uchar* top = sptr + srcstep*MAX(y-wsz2-1,0);
138         const uchar* bottom = sptr + srcstep*MIN(y+wsz2,size.height-1);
139         const uchar* prev = sptr + srcstep*MAX(y-1,0);
140         const uchar* curr = sptr + srcstep*y;
141         const uchar* next = sptr + srcstep*MIN(y+1,size.height-1);
142         uchar* dptr = dst.ptr<uchar>(y);
143 
144         for( x = 0; x < size.width; x++ )
145             vsum[x] = (ushort)(vsum[x] + bottom[x] - top[x]);
146 
147         for( x = 0; x <= wsz2; x++ )
148         {
149             vsum[-x-1] = vsum[0];
150             vsum[size.width+x] = vsum[size.width-1];
151         }
152 
153         int sum = vsum[0]*(wsz2 + 1);
154         for( x = 1; x <= wsz2; x++ )
155             sum += vsum[x];
156 
157         int val = ((curr[0]*5 + curr[1] + prev[0] + next[0])*scale_g - sum*scale_s) >> 10;
158         dptr[0] = tab[val + OFS];
159 
160         for( x = 1; x < size.width-1; x++ )
161         {
162             sum += vsum[x+wsz2] - vsum[x-wsz2-1];
163             val = ((curr[x]*4 + curr[x-1] + curr[x+1] + prev[x] + next[x])*scale_g - sum*scale_s) >> 10;
164             dptr[x] = tab[val + OFS];
165         }
166 
167         sum += vsum[x+wsz2] - vsum[x-wsz2-1];
168         val = ((curr[x]*5 + curr[x-1] + prev[x] + next[x])*scale_g - sum*scale_s) >> 10;
169         dptr[x] = tab[val + OFS];
170     }
171 }
172 
ocl_prefilter_xsobel(InputArray _input,OutputArray _output,int prefilterCap)173 static bool ocl_prefilter_xsobel(InputArray _input, OutputArray _output, int prefilterCap)
174 {
175     ocl::Kernel k("prefilter_xsobel", ocl::calib3d::stereobm_oclsrc);
176     if(k.empty())
177         return false;
178 
179     UMat input = _input.getUMat(), output;
180     _output.create(input.size(), input.type());
181     output = _output.getUMat();
182 
183     size_t globalThreads[3] = { input.cols, input.rows, 1 };
184 
185     k.args(ocl::KernelArg::PtrReadOnly(input), ocl::KernelArg::PtrWriteOnly(output), input.rows, input.cols, prefilterCap);
186 
187     return k.run(2, globalThreads, NULL, false);
188 }
189 
190 static void
prefilterXSobel(const Mat & src,Mat & dst,int ftzero)191 prefilterXSobel( const Mat& src, Mat& dst, int ftzero )
192 {
193     int x, y;
194     const int OFS = 256*4, TABSZ = OFS*2 + 256;
195     uchar tab[TABSZ];
196     Size size = src.size();
197 
198     for( x = 0; x < TABSZ; x++ )
199         tab[x] = (uchar)(x - OFS < -ftzero ? 0 : x - OFS > ftzero ? ftzero*2 : x - OFS + ftzero);
200     uchar val0 = tab[0 + OFS];
201 
202 #if CV_SSE2
203     volatile bool useSIMD = checkHardwareSupport(CV_CPU_SSE2);
204 #endif
205 
206     for( y = 0; y < size.height-1; y += 2 )
207     {
208         const uchar* srow1 = src.ptr<uchar>(y);
209         const uchar* srow0 = y > 0 ? srow1 - src.step : size.height > 1 ? srow1 + src.step : srow1;
210         const uchar* srow2 = y < size.height-1 ? srow1 + src.step : size.height > 1 ? srow1 - src.step : srow1;
211         const uchar* srow3 = y < size.height-2 ? srow1 + src.step*2 : srow1;
212         uchar* dptr0 = dst.ptr<uchar>(y);
213         uchar* dptr1 = dptr0 + dst.step;
214 
215         dptr0[0] = dptr0[size.width-1] = dptr1[0] = dptr1[size.width-1] = val0;
216         x = 1;
217 
218 #if CV_NEON
219         int16x8_t ftz = vdupq_n_s16 ((short) ftzero);
220         uint8x8_t ftz2 = vdup_n_u8 (cv::saturate_cast<uchar>(ftzero*2));
221 
222         for(; x <=size.width-9; x += 8 )
223         {
224             uint8x8_t c0 = vld1_u8 (srow0 + x - 1);
225             uint8x8_t c1 = vld1_u8 (srow1 + x - 1);
226             uint8x8_t d0 = vld1_u8 (srow0 + x + 1);
227             uint8x8_t d1 = vld1_u8 (srow1 + x + 1);
228 
229             int16x8_t t0 = vreinterpretq_s16_u16 (vsubl_u8 (d0, c0));
230             int16x8_t t1 = vreinterpretq_s16_u16 (vsubl_u8 (d1, c1));
231 
232             uint8x8_t c2 = vld1_u8 (srow2 + x - 1);
233             uint8x8_t c3 = vld1_u8 (srow3 + x - 1);
234             uint8x8_t d2 = vld1_u8 (srow2 + x + 1);
235             uint8x8_t d3 = vld1_u8 (srow3 + x + 1);
236 
237             int16x8_t t2 = vreinterpretq_s16_u16 (vsubl_u8 (d2, c2));
238             int16x8_t t3 = vreinterpretq_s16_u16 (vsubl_u8 (d3, c3));
239 
240             int16x8_t v0 = vaddq_s16 (vaddq_s16 (t2, t0), vaddq_s16 (t1, t1));
241             int16x8_t v1 = vaddq_s16 (vaddq_s16 (t3, t1), vaddq_s16 (t2, t2));
242 
243 
244             uint8x8_t v0_u8 = vqmovun_s16 (vaddq_s16 (v0, ftz));
245             uint8x8_t v1_u8 = vqmovun_s16 (vaddq_s16 (v1, ftz));
246             v0_u8 =  vmin_u8 (v0_u8, ftz2);
247             v1_u8 =  vmin_u8 (v1_u8, ftz2);
248             vqmovun_s16 (vaddq_s16 (v1, ftz));
249 
250             vst1_u8 (dptr0 + x, v0_u8);
251             vst1_u8 (dptr1 + x, v1_u8);
252         }
253 #elif CV_SSE2
254         if( useSIMD )
255         {
256             __m128i z = _mm_setzero_si128(), ftz = _mm_set1_epi16((short)ftzero),
257             ftz2 = _mm_set1_epi8(cv::saturate_cast<uchar>(ftzero*2));
258             for( ; x <= size.width-9; x += 8 )
259             {
260                 __m128i c0 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i*)(srow0 + x - 1)), z);
261                 __m128i c1 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i*)(srow1 + x - 1)), z);
262                 __m128i d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i*)(srow0 + x + 1)), z);
263                 __m128i d1 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i*)(srow1 + x + 1)), z);
264 
265                 d0 = _mm_sub_epi16(d0, c0);
266                 d1 = _mm_sub_epi16(d1, c1);
267 
268                 __m128i c2 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i*)(srow2 + x - 1)), z);
269                 __m128i c3 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i*)(srow3 + x - 1)), z);
270                 __m128i d2 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i*)(srow2 + x + 1)), z);
271                 __m128i d3 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i*)(srow3 + x + 1)), z);
272 
273                 d2 = _mm_sub_epi16(d2, c2);
274                 d3 = _mm_sub_epi16(d3, c3);
275 
276                 __m128i v0 = _mm_add_epi16(d0, _mm_add_epi16(d2, _mm_add_epi16(d1, d1)));
277                 __m128i v1 = _mm_add_epi16(d1, _mm_add_epi16(d3, _mm_add_epi16(d2, d2)));
278                 v0 = _mm_packus_epi16(_mm_add_epi16(v0, ftz), _mm_add_epi16(v1, ftz));
279                 v0 = _mm_min_epu8(v0, ftz2);
280 
281                 _mm_storel_epi64((__m128i*)(dptr0 + x), v0);
282                 _mm_storel_epi64((__m128i*)(dptr1 + x), _mm_unpackhi_epi64(v0, v0));
283             }
284         }
285 #endif
286 
287         for( ; x < size.width-1; x++ )
288         {
289             int d0 = srow0[x+1] - srow0[x-1], d1 = srow1[x+1] - srow1[x-1],
290             d2 = srow2[x+1] - srow2[x-1], d3 = srow3[x+1] - srow3[x-1];
291             int v0 = tab[d0 + d1*2 + d2 + OFS];
292             int v1 = tab[d1 + d2*2 + d3 + OFS];
293             dptr0[x] = (uchar)v0;
294             dptr1[x] = (uchar)v1;
295         }
296     }
297 
298 #if CV_NEON
299     uint8x16_t val0_16 = vdupq_n_u8 (val0);
300 #endif
301 
302     for( ; y < size.height; y++ )
303     {
304         uchar* dptr = dst.ptr<uchar>(y);
305         x = 0;
306     #if CV_NEON
307         for(; x <= size.width-16; x+=16 )
308             vst1q_u8 (dptr + x, val0_16);
309     #endif
310         for(; x < size.width; x++ )
311             dptr[x] = val0;
312     }
313 }
314 
315 
316 static const int DISPARITY_SHIFT = 4;
317 
318 #if CV_SSE2
findStereoCorrespondenceBM_SSE2(const Mat & left,const Mat & right,Mat & disp,Mat & cost,StereoBMParams & state,uchar * buf,int _dy0,int _dy1)319 static void findStereoCorrespondenceBM_SSE2( const Mat& left, const Mat& right,
320                                             Mat& disp, Mat& cost, StereoBMParams& state,
321                                             uchar* buf, int _dy0, int _dy1 )
322 {
323     const int ALIGN = 16;
324     int x, y, d;
325     int wsz = state.SADWindowSize, wsz2 = wsz/2;
326     int dy0 = MIN(_dy0, wsz2+1), dy1 = MIN(_dy1, wsz2+1);
327     int ndisp = state.numDisparities;
328     int mindisp = state.minDisparity;
329     int lofs = MAX(ndisp - 1 + mindisp, 0);
330     int rofs = -MIN(ndisp - 1 + mindisp, 0);
331     int width = left.cols, height = left.rows;
332     int width1 = width - rofs - ndisp + 1;
333     int ftzero = state.preFilterCap;
334     int textureThreshold = state.textureThreshold;
335     int uniquenessRatio = state.uniquenessRatio;
336     short FILTERED = (short)((mindisp - 1) << DISPARITY_SHIFT);
337 
338     ushort *sad, *hsad0, *hsad, *hsad_sub;
339     int *htext;
340     uchar *cbuf0, *cbuf;
341     const uchar* lptr0 = left.ptr() + lofs;
342     const uchar* rptr0 = right.ptr() + rofs;
343     const uchar *lptr, *lptr_sub, *rptr;
344     short* dptr = disp.ptr<short>();
345     int sstep = (int)left.step;
346     int dstep = (int)(disp.step/sizeof(dptr[0]));
347     int cstep = (height + dy0 + dy1)*ndisp;
348     short costbuf = 0;
349     int coststep = cost.data ? (int)(cost.step/sizeof(costbuf)) : 0;
350     const int TABSZ = 256;
351     uchar tab[TABSZ];
352     const __m128i d0_8 = _mm_setr_epi16(0,1,2,3,4,5,6,7), dd_8 = _mm_set1_epi16(8);
353 
354     sad = (ushort*)alignPtr(buf + sizeof(sad[0]), ALIGN);
355     hsad0 = (ushort*)alignPtr(sad + ndisp + 1 + dy0*ndisp, ALIGN);
356     htext = (int*)alignPtr((int*)(hsad0 + (height+dy1)*ndisp) + wsz2 + 2, ALIGN);
357     cbuf0 = (uchar*)alignPtr((uchar*)(htext + height + wsz2 + 2) + dy0*ndisp, ALIGN);
358 
359     for( x = 0; x < TABSZ; x++ )
360         tab[x] = (uchar)std::abs(x - ftzero);
361 
362     // initialize buffers
363     memset( hsad0 - dy0*ndisp, 0, (height + dy0 + dy1)*ndisp*sizeof(hsad0[0]) );
364     memset( htext - wsz2 - 1, 0, (height + wsz + 1)*sizeof(htext[0]) );
365 
366     for( x = -wsz2-1; x < wsz2; x++ )
367     {
368         hsad = hsad0 - dy0*ndisp; cbuf = cbuf0 + (x + wsz2 + 1)*cstep - dy0*ndisp;
369         lptr = lptr0 + MIN(MAX(x, -lofs), width-lofs-1) - dy0*sstep;
370         rptr = rptr0 + MIN(MAX(x, -rofs), width-rofs-1) - dy0*sstep;
371 
372         for( y = -dy0; y < height + dy1; y++, hsad += ndisp, cbuf += ndisp, lptr += sstep, rptr += sstep )
373         {
374             int lval = lptr[0];
375             __m128i lv = _mm_set1_epi8((char)lval), z = _mm_setzero_si128();
376             for( d = 0; d < ndisp; d += 16 )
377             {
378                 __m128i rv = _mm_loadu_si128((const __m128i*)(rptr + d));
379                 __m128i hsad_l = _mm_load_si128((__m128i*)(hsad + d));
380                 __m128i hsad_h = _mm_load_si128((__m128i*)(hsad + d + 8));
381                 __m128i diff = _mm_adds_epu8(_mm_subs_epu8(lv, rv), _mm_subs_epu8(rv, lv));
382                 _mm_store_si128((__m128i*)(cbuf + d), diff);
383                 hsad_l = _mm_add_epi16(hsad_l, _mm_unpacklo_epi8(diff,z));
384                 hsad_h = _mm_add_epi16(hsad_h, _mm_unpackhi_epi8(diff,z));
385                 _mm_store_si128((__m128i*)(hsad + d), hsad_l);
386                 _mm_store_si128((__m128i*)(hsad + d + 8), hsad_h);
387             }
388             htext[y] += tab[lval];
389         }
390     }
391 
392     // initialize the left and right borders of the disparity map
393     for( y = 0; y < height; y++ )
394     {
395         for( x = 0; x < lofs; x++ )
396             dptr[y*dstep + x] = FILTERED;
397         for( x = lofs + width1; x < width; x++ )
398             dptr[y*dstep + x] = FILTERED;
399     }
400     dptr += lofs;
401 
402     for( x = 0; x < width1; x++, dptr++ )
403     {
404         short* costptr = cost.data ? cost.ptr<short>() + lofs + x : &costbuf;
405         int x0 = x - wsz2 - 1, x1 = x + wsz2;
406         const uchar* cbuf_sub = cbuf0 + ((x0 + wsz2 + 1) % (wsz + 1))*cstep - dy0*ndisp;
407         cbuf = cbuf0 + ((x1 + wsz2 + 1) % (wsz + 1))*cstep - dy0*ndisp;
408         hsad = hsad0 - dy0*ndisp;
409         lptr_sub = lptr0 + MIN(MAX(x0, -lofs), width-1-lofs) - dy0*sstep;
410         lptr = lptr0 + MIN(MAX(x1, -lofs), width-1-lofs) - dy0*sstep;
411         rptr = rptr0 + MIN(MAX(x1, -rofs), width-1-rofs) - dy0*sstep;
412 
413         for( y = -dy0; y < height + dy1; y++, cbuf += ndisp, cbuf_sub += ndisp,
414             hsad += ndisp, lptr += sstep, lptr_sub += sstep, rptr += sstep )
415         {
416             int lval = lptr[0];
417             __m128i lv = _mm_set1_epi8((char)lval), z = _mm_setzero_si128();
418             for( d = 0; d < ndisp; d += 16 )
419             {
420                 __m128i rv = _mm_loadu_si128((const __m128i*)(rptr + d));
421                 __m128i hsad_l = _mm_load_si128((__m128i*)(hsad + d));
422                 __m128i hsad_h = _mm_load_si128((__m128i*)(hsad + d + 8));
423                 __m128i cbs = _mm_load_si128((const __m128i*)(cbuf_sub + d));
424                 __m128i diff = _mm_adds_epu8(_mm_subs_epu8(lv, rv), _mm_subs_epu8(rv, lv));
425                 __m128i diff_h = _mm_sub_epi16(_mm_unpackhi_epi8(diff, z), _mm_unpackhi_epi8(cbs, z));
426                 _mm_store_si128((__m128i*)(cbuf + d), diff);
427                 diff = _mm_sub_epi16(_mm_unpacklo_epi8(diff, z), _mm_unpacklo_epi8(cbs, z));
428                 hsad_h = _mm_add_epi16(hsad_h, diff_h);
429                 hsad_l = _mm_add_epi16(hsad_l, diff);
430                 _mm_store_si128((__m128i*)(hsad + d), hsad_l);
431                 _mm_store_si128((__m128i*)(hsad + d + 8), hsad_h);
432             }
433             htext[y] += tab[lval] - tab[lptr_sub[0]];
434         }
435 
436         // fill borders
437         for( y = dy1; y <= wsz2; y++ )
438             htext[height+y] = htext[height+dy1-1];
439         for( y = -wsz2-1; y < -dy0; y++ )
440             htext[y] = htext[-dy0];
441 
442         // initialize sums
443         for( d = 0; d < ndisp; d++ )
444             sad[d] = (ushort)(hsad0[d-ndisp*dy0]*(wsz2 + 2 - dy0));
445 
446         hsad = hsad0 + (1 - dy0)*ndisp;
447         for( y = 1 - dy0; y < wsz2; y++, hsad += ndisp )
448             for( d = 0; d < ndisp; d += 16 )
449             {
450                 __m128i s0 = _mm_load_si128((__m128i*)(sad + d));
451                 __m128i s1 = _mm_load_si128((__m128i*)(sad + d + 8));
452                 __m128i t0 = _mm_load_si128((__m128i*)(hsad + d));
453                 __m128i t1 = _mm_load_si128((__m128i*)(hsad + d + 8));
454                 s0 = _mm_add_epi16(s0, t0);
455                 s1 = _mm_add_epi16(s1, t1);
456                 _mm_store_si128((__m128i*)(sad + d), s0);
457                 _mm_store_si128((__m128i*)(sad + d + 8), s1);
458             }
459         int tsum = 0;
460         for( y = -wsz2-1; y < wsz2; y++ )
461             tsum += htext[y];
462 
463         // finally, start the real processing
464         for( y = 0; y < height; y++ )
465         {
466             int minsad = INT_MAX, mind = -1;
467             hsad = hsad0 + MIN(y + wsz2, height+dy1-1)*ndisp;
468             hsad_sub = hsad0 + MAX(y - wsz2 - 1, -dy0)*ndisp;
469             __m128i minsad8 = _mm_set1_epi16(SHRT_MAX);
470             __m128i mind8 = _mm_set1_epi16(0), d8 = d0_8, mask;
471 
472             for( d = 0; d < ndisp; d += 16 )
473             {
474                 __m128i u0 = _mm_load_si128((__m128i*)(hsad_sub + d));
475                 __m128i u1 = _mm_load_si128((__m128i*)(hsad + d));
476 
477                 __m128i v0 = _mm_load_si128((__m128i*)(hsad_sub + d + 8));
478                 __m128i v1 = _mm_load_si128((__m128i*)(hsad + d + 8));
479 
480                 __m128i usad8 = _mm_load_si128((__m128i*)(sad + d));
481                 __m128i vsad8 = _mm_load_si128((__m128i*)(sad + d + 8));
482 
483                 u1 = _mm_sub_epi16(u1, u0);
484                 v1 = _mm_sub_epi16(v1, v0);
485                 usad8 = _mm_add_epi16(usad8, u1);
486                 vsad8 = _mm_add_epi16(vsad8, v1);
487 
488                 mask = _mm_cmpgt_epi16(minsad8, usad8);
489                 minsad8 = _mm_min_epi16(minsad8, usad8);
490                 mind8 = _mm_max_epi16(mind8, _mm_and_si128(mask, d8));
491 
492                 _mm_store_si128((__m128i*)(sad + d), usad8);
493                 _mm_store_si128((__m128i*)(sad + d + 8), vsad8);
494 
495                 mask = _mm_cmpgt_epi16(minsad8, vsad8);
496                 minsad8 = _mm_min_epi16(minsad8, vsad8);
497 
498                 d8 = _mm_add_epi16(d8, dd_8);
499                 mind8 = _mm_max_epi16(mind8, _mm_and_si128(mask, d8));
500                 d8 = _mm_add_epi16(d8, dd_8);
501             }
502 
503             tsum += htext[y + wsz2] - htext[y - wsz2 - 1];
504             if( tsum < textureThreshold )
505             {
506                 dptr[y*dstep] = FILTERED;
507                 continue;
508             }
509 
510             ushort CV_DECL_ALIGNED(16) minsad_buf[8], mind_buf[8];
511             _mm_store_si128((__m128i*)minsad_buf, minsad8);
512             _mm_store_si128((__m128i*)mind_buf, mind8);
513             for( d = 0; d < 8; d++ )
514                 if(minsad > (int)minsad_buf[d] || (minsad == (int)minsad_buf[d] && mind > mind_buf[d]))
515                 {
516                     minsad = minsad_buf[d];
517                     mind = mind_buf[d];
518                 }
519 
520             if( uniquenessRatio > 0 )
521             {
522                 int thresh = minsad + (minsad * uniquenessRatio/100);
523                 __m128i thresh8 = _mm_set1_epi16((short)(thresh + 1));
524                 __m128i d1 = _mm_set1_epi16((short)(mind-1)), d2 = _mm_set1_epi16((short)(mind+1));
525                 __m128i dd_16 = _mm_add_epi16(dd_8, dd_8);
526                 d8 = _mm_sub_epi16(d0_8, dd_16);
527 
528                 for( d = 0; d < ndisp; d += 16 )
529                 {
530                     __m128i usad8 = _mm_load_si128((__m128i*)(sad + d));
531                     __m128i vsad8 = _mm_load_si128((__m128i*)(sad + d + 8));
532                     mask = _mm_cmpgt_epi16( thresh8, _mm_min_epi16(usad8,vsad8));
533                     d8 = _mm_add_epi16(d8, dd_16);
534                     if( !_mm_movemask_epi8(mask) )
535                         continue;
536                     mask = _mm_cmpgt_epi16( thresh8, usad8);
537                     mask = _mm_and_si128(mask, _mm_or_si128(_mm_cmpgt_epi16(d1,d8), _mm_cmpgt_epi16(d8,d2)));
538                     if( _mm_movemask_epi8(mask) )
539                         break;
540                     __m128i t8 = _mm_add_epi16(d8, dd_8);
541                     mask = _mm_cmpgt_epi16( thresh8, vsad8);
542                     mask = _mm_and_si128(mask, _mm_or_si128(_mm_cmpgt_epi16(d1,t8), _mm_cmpgt_epi16(t8,d2)));
543                     if( _mm_movemask_epi8(mask) )
544                         break;
545                 }
546                 if( d < ndisp )
547                 {
548                     dptr[y*dstep] = FILTERED;
549                     continue;
550                 }
551             }
552 
553             if( 0 < mind && mind < ndisp - 1 )
554             {
555                 int p = sad[mind+1], n = sad[mind-1];
556                 d = p + n - 2*sad[mind] + std::abs(p - n);
557                 dptr[y*dstep] = (short)(((ndisp - mind - 1 + mindisp)*256 + (d != 0 ? (p-n)*256/d : 0) + 15) >> 4);
558             }
559             else
560                 dptr[y*dstep] = (short)((ndisp - mind - 1 + mindisp)*16);
561             costptr[y*coststep] = sad[mind];
562         }
563     }
564 }
565 #endif
566 
567 static void
findStereoCorrespondenceBM(const Mat & left,const Mat & right,Mat & disp,Mat & cost,const StereoBMParams & state,uchar * buf,int _dy0,int _dy1)568 findStereoCorrespondenceBM( const Mat& left, const Mat& right,
569                            Mat& disp, Mat& cost, const StereoBMParams& state,
570                            uchar* buf, int _dy0, int _dy1 )
571 {
572 
573     const int ALIGN = 16;
574     int x, y, d;
575     int wsz = state.SADWindowSize, wsz2 = wsz/2;
576     int dy0 = MIN(_dy0, wsz2+1), dy1 = MIN(_dy1, wsz2+1);
577     int ndisp = state.numDisparities;
578     int mindisp = state.minDisparity;
579     int lofs = MAX(ndisp - 1 + mindisp, 0);
580     int rofs = -MIN(ndisp - 1 + mindisp, 0);
581     int width = left.cols, height = left.rows;
582     int width1 = width - rofs - ndisp + 1;
583     int ftzero = state.preFilterCap;
584     int textureThreshold = state.textureThreshold;
585     int uniquenessRatio = state.uniquenessRatio;
586     short FILTERED = (short)((mindisp - 1) << DISPARITY_SHIFT);
587 
588 #if CV_NEON
589     CV_Assert (ndisp % 8 == 0);
590     int32_t d0_4_temp [4];
591     for (int i = 0; i < 4; i ++)
592         d0_4_temp[i] = i;
593     int32x4_t d0_4 = vld1q_s32 (d0_4_temp);
594     int32x4_t dd_4 = vdupq_n_s32 (4);
595 #endif
596 
597     int *sad, *hsad0, *hsad, *hsad_sub, *htext;
598     uchar *cbuf0, *cbuf;
599     const uchar* lptr0 = left.ptr() + lofs;
600     const uchar* rptr0 = right.ptr() + rofs;
601     const uchar *lptr, *lptr_sub, *rptr;
602     short* dptr = disp.ptr<short>();
603     int sstep = (int)left.step;
604     int dstep = (int)(disp.step/sizeof(dptr[0]));
605     int cstep = (height+dy0+dy1)*ndisp;
606     int costbuf = 0;
607     int coststep = cost.data ? (int)(cost.step/sizeof(costbuf)) : 0;
608     const int TABSZ = 256;
609     uchar tab[TABSZ];
610 
611     sad = (int*)alignPtr(buf + sizeof(sad[0]), ALIGN);
612     hsad0 = (int*)alignPtr(sad + ndisp + 1 + dy0*ndisp, ALIGN);
613     htext = (int*)alignPtr((int*)(hsad0 + (height+dy1)*ndisp) + wsz2 + 2, ALIGN);
614     cbuf0 = (uchar*)alignPtr((uchar*)(htext + height + wsz2 + 2) + dy0*ndisp, ALIGN);
615 
616     for( x = 0; x < TABSZ; x++ )
617         tab[x] = (uchar)std::abs(x - ftzero);
618 
619     // initialize buffers
620     memset( hsad0 - dy0*ndisp, 0, (height + dy0 + dy1)*ndisp*sizeof(hsad0[0]) );
621     memset( htext - wsz2 - 1, 0, (height + wsz + 1)*sizeof(htext[0]) );
622 
623     for( x = -wsz2-1; x < wsz2; x++ )
624     {
625         hsad = hsad0 - dy0*ndisp; cbuf = cbuf0 + (x + wsz2 + 1)*cstep - dy0*ndisp;
626         lptr = lptr0 + std::min(std::max(x, -lofs), width-lofs-1) - dy0*sstep;
627         rptr = rptr0 + std::min(std::max(x, -rofs), width-rofs-1) - dy0*sstep;
628         for( y = -dy0; y < height + dy1; y++, hsad += ndisp, cbuf += ndisp, lptr += sstep, rptr += sstep )
629         {
630             int lval = lptr[0];
631         #if CV_NEON
632             int16x8_t lv = vdupq_n_s16 ((int16_t)lval);
633 
634             for( d = 0; d < ndisp; d += 8 )
635             {
636                 int16x8_t rv = vreinterpretq_s16_u16 (vmovl_u8 (vld1_u8 (rptr + d)));
637                 int32x4_t hsad_l = vld1q_s32 (hsad + d);
638                 int32x4_t hsad_h = vld1q_s32 (hsad + d + 4);
639                 int16x8_t diff = vabdq_s16 (lv, rv);
640                 vst1_u8 (cbuf + d, vmovn_u16(vreinterpretq_u16_s16(diff)));
641                 hsad_l = vaddq_s32 (hsad_l, vmovl_s16(vget_low_s16 (diff)));
642                 hsad_h = vaddq_s32 (hsad_h, vmovl_s16(vget_high_s16 (diff)));
643                 vst1q_s32 ((hsad + d), hsad_l);
644                 vst1q_s32 ((hsad + d + 4), hsad_h);
645             }
646         #else
647             for( d = 0; d < ndisp; d++ )
648             {
649                 int diff = std::abs(lval - rptr[d]);
650                 cbuf[d] = (uchar)diff;
651                 hsad[d] = (int)(hsad[d] + diff);
652             }
653         #endif
654             htext[y] += tab[lval];
655         }
656     }
657 
658     // initialize the left and right borders of the disparity map
659     for( y = 0; y < height; y++ )
660     {
661         for( x = 0; x < lofs; x++ )
662             dptr[y*dstep + x] = FILTERED;
663         for( x = lofs + width1; x < width; x++ )
664             dptr[y*dstep + x] = FILTERED;
665     }
666     dptr += lofs;
667 
668     for( x = 0; x < width1; x++, dptr++ )
669     {
670         int* costptr = cost.data ? cost.ptr<int>() + lofs + x : &costbuf;
671         int x0 = x - wsz2 - 1, x1 = x + wsz2;
672         const uchar* cbuf_sub = cbuf0 + ((x0 + wsz2 + 1) % (wsz + 1))*cstep - dy0*ndisp;
673         cbuf = cbuf0 + ((x1 + wsz2 + 1) % (wsz + 1))*cstep - dy0*ndisp;
674         hsad = hsad0 - dy0*ndisp;
675         lptr_sub = lptr0 + MIN(MAX(x0, -lofs), width-1-lofs) - dy0*sstep;
676         lptr = lptr0 + MIN(MAX(x1, -lofs), width-1-lofs) - dy0*sstep;
677         rptr = rptr0 + MIN(MAX(x1, -rofs), width-1-rofs) - dy0*sstep;
678 
679         for( y = -dy0; y < height + dy1; y++, cbuf += ndisp, cbuf_sub += ndisp,
680             hsad += ndisp, lptr += sstep, lptr_sub += sstep, rptr += sstep )
681         {
682             int lval = lptr[0];
683         #if CV_NEON
684             int16x8_t lv = vdupq_n_s16 ((int16_t)lval);
685             for( d = 0; d < ndisp; d += 8 )
686             {
687                 int16x8_t rv = vreinterpretq_s16_u16 (vmovl_u8 (vld1_u8 (rptr + d)));
688                 int32x4_t hsad_l = vld1q_s32 (hsad + d);
689                 int32x4_t hsad_h = vld1q_s32 (hsad + d + 4);
690                 int16x8_t cbs = vreinterpretq_s16_u16 (vmovl_u8 (vld1_u8 (cbuf_sub + d)));
691                 int16x8_t diff = vabdq_s16 (lv, rv);
692                 int32x4_t diff_h = vsubl_s16 (vget_high_s16 (diff), vget_high_s16 (cbs));
693                 int32x4_t diff_l = vsubl_s16 (vget_low_s16 (diff), vget_low_s16 (cbs));
694                 vst1_u8 (cbuf + d, vmovn_u16(vreinterpretq_u16_s16(diff)));
695                 hsad_h = vaddq_s32 (hsad_h, diff_h);
696                 hsad_l = vaddq_s32 (hsad_l, diff_l);
697                 vst1q_s32 ((hsad + d), hsad_l);
698                 vst1q_s32 ((hsad + d + 4), hsad_h);
699             }
700         #else
701             for( d = 0; d < ndisp; d++ )
702             {
703                 int diff = std::abs(lval - rptr[d]);
704                 cbuf[d] = (uchar)diff;
705                 hsad[d] = hsad[d] + diff - cbuf_sub[d];
706             }
707         #endif
708             htext[y] += tab[lval] - tab[lptr_sub[0]];
709         }
710 
711         // fill borders
712         for( y = dy1; y <= wsz2; y++ )
713             htext[height+y] = htext[height+dy1-1];
714         for( y = -wsz2-1; y < -dy0; y++ )
715             htext[y] = htext[-dy0];
716 
717         // initialize sums
718         for( d = 0; d < ndisp; d++ )
719             sad[d] = (int)(hsad0[d-ndisp*dy0]*(wsz2 + 2 - dy0));
720 
721         hsad = hsad0 + (1 - dy0)*ndisp;
722         for( y = 1 - dy0; y < wsz2; y++, hsad += ndisp )
723         {
724         #if CV_NEON
725             for( d = 0; d <= ndisp-8; d += 8 )
726             {
727                 int32x4_t s0 = vld1q_s32 (sad + d);
728                 int32x4_t s1 = vld1q_s32 (sad + d + 4);
729                 int32x4_t t0 = vld1q_s32 (hsad + d);
730                 int32x4_t t1 = vld1q_s32 (hsad + d + 4);
731                 s0 = vaddq_s32 (s0, t0);
732                 s1 = vaddq_s32 (s1, t1);
733                 vst1q_s32 (sad + d, s0);
734                 vst1q_s32 (sad + d + 4, s1);
735             }
736         #else
737             for( d = 0; d < ndisp; d++ )
738                 sad[d] = (int)(sad[d] + hsad[d]);
739         #endif
740         }
741         int tsum = 0;
742         for( y = -wsz2-1; y < wsz2; y++ )
743             tsum += htext[y];
744 
745         // finally, start the real processing
746         for( y = 0; y < height; y++ )
747         {
748             int minsad = INT_MAX, mind = -1;
749             hsad = hsad0 + MIN(y + wsz2, height+dy1-1)*ndisp;
750             hsad_sub = hsad0 + MAX(y - wsz2 - 1, -dy0)*ndisp;
751         #if CV_NEON
752             int32x4_t minsad4 = vdupq_n_s32 (INT_MAX);
753             int32x4_t mind4 = vdupq_n_s32(0), d4 = d0_4;
754 
755             for( d = 0; d <= ndisp-8; d += 8 )
756             {
757                 int32x4_t u0 = vld1q_s32 (hsad_sub + d);
758                 int32x4_t u1 = vld1q_s32 (hsad + d);
759 
760                 int32x4_t v0 = vld1q_s32 (hsad_sub + d + 4);
761                 int32x4_t v1 = vld1q_s32 (hsad + d + 4);
762 
763                 int32x4_t usad4 = vld1q_s32(sad + d);
764                 int32x4_t vsad4 = vld1q_s32(sad + d + 4);
765 
766                 u1 = vsubq_s32 (u1, u0);
767                 v1 = vsubq_s32 (v1, v0);
768                 usad4 = vaddq_s32 (usad4, u1);
769                 vsad4 = vaddq_s32 (vsad4, v1);
770 
771                 uint32x4_t mask = vcgtq_s32 (minsad4, usad4);
772                 minsad4 = vminq_s32 (minsad4, usad4);
773                 mind4 = vbslq_s32(mask, d4, mind4);
774 
775                 vst1q_s32 (sad + d, usad4);
776                 vst1q_s32 (sad + d + 4, vsad4);
777                 d4 = vaddq_s32 (d4, dd_4);
778 
779                 mask = vcgtq_s32 (minsad4, vsad4);
780                 minsad4 = vminq_s32 (minsad4, vsad4);
781                 mind4 = vbslq_s32(mask, d4, mind4);
782 
783                 d4 = vaddq_s32 (d4, dd_4);
784 
785             }
786             int32x2_t mind4_h = vget_high_s32 (mind4);
787             int32x2_t mind4_l = vget_low_s32 (mind4);
788             int32x2_t minsad4_h = vget_high_s32 (minsad4);
789             int32x2_t minsad4_l = vget_low_s32 (minsad4);
790 
791             uint32x2_t mask = vorr_u32 (vclt_s32 (minsad4_h, minsad4_l), vand_u32 (vceq_s32 (minsad4_h, minsad4_l), vclt_s32 (mind4_h, mind4_l)));
792             mind4_h = vbsl_s32 (mask, mind4_h, mind4_l);
793             minsad4_h = vbsl_s32 (mask, minsad4_h, minsad4_l);
794 
795             mind4_l = vext_s32 (mind4_h,mind4_h,1);
796             minsad4_l = vext_s32 (minsad4_h,minsad4_h,1);
797 
798             mask = vorr_u32 (vclt_s32 (minsad4_h, minsad4_l), vand_u32 (vceq_s32 (minsad4_h, minsad4_l), vclt_s32 (mind4_h, mind4_l)));
799             mind4_h = vbsl_s32 (mask, mind4_h, mind4_l);
800             minsad4_h = vbsl_s32 (mask, minsad4_h, minsad4_l);
801 
802             mind = (int) vget_lane_s32 (mind4_h, 0);
803             minsad = sad[mind];
804 
805         #else
806             for( d = 0; d < ndisp; d++ )
807             {
808                 int currsad = sad[d] + hsad[d] - hsad_sub[d];
809                 sad[d] = currsad;
810                 if( currsad < minsad )
811                 {
812                     minsad = currsad;
813                     mind = d;
814                 }
815             }
816         #endif
817 
818             tsum += htext[y + wsz2] - htext[y - wsz2 - 1];
819             if( tsum < textureThreshold )
820             {
821                 dptr[y*dstep] = FILTERED;
822                 continue;
823             }
824 
825             if( uniquenessRatio > 0 )
826             {
827                 int thresh = minsad + (minsad * uniquenessRatio/100);
828                 for( d = 0; d < ndisp; d++ )
829                 {
830                     if( (d < mind-1 || d > mind+1) && sad[d] <= thresh)
831                         break;
832                 }
833                 if( d < ndisp )
834                 {
835                     dptr[y*dstep] = FILTERED;
836                     continue;
837                 }
838             }
839 
840             {
841                 sad[-1] = sad[1];
842                 sad[ndisp] = sad[ndisp-2];
843                 int p = sad[mind+1], n = sad[mind-1];
844                 d = p + n - 2*sad[mind] + std::abs(p - n);
845                 dptr[y*dstep] = (short)(((ndisp - mind - 1 + mindisp)*256 + (d != 0 ? (p-n)*256/d : 0) + 15) >> 4);
846                 costptr[y*coststep] = sad[mind];
847             }
848         }
849     }
850 }
851 
ocl_prefiltering(InputArray left0,InputArray right0,OutputArray left,OutputArray right,StereoBMParams * state)852 static bool ocl_prefiltering(InputArray left0, InputArray right0, OutputArray left, OutputArray right, StereoBMParams* state)
853 {
854     if( state->preFilterType == StereoBM::PREFILTER_NORMALIZED_RESPONSE )
855     {
856         if(!ocl_prefilter_norm( left0, left, state->preFilterSize, state->preFilterCap))
857             return false;
858         if(!ocl_prefilter_norm( right0, right, state->preFilterSize, state->preFilterCap))
859             return false;
860     }
861     else
862     {
863         if(!ocl_prefilter_xsobel( left0, left, state->preFilterCap ))
864             return false;
865         if(!ocl_prefilter_xsobel( right0, right, state->preFilterCap))
866             return false;
867     }
868     return true;
869 }
870 
871 struct PrefilterInvoker : public ParallelLoopBody
872 {
PrefilterInvokercv::PrefilterInvoker873     PrefilterInvoker(const Mat& left0, const Mat& right0, Mat& left, Mat& right,
874                      uchar* buf0, uchar* buf1, StereoBMParams* _state)
875     {
876         imgs0[0] = &left0; imgs0[1] = &right0;
877         imgs[0] = &left; imgs[1] = &right;
878         buf[0] = buf0; buf[1] = buf1;
879         state = _state;
880     }
881 
operator ()cv::PrefilterInvoker882     void operator()( const Range& range ) const
883     {
884         for( int i = range.start; i < range.end; i++ )
885         {
886             if( state->preFilterType == StereoBM::PREFILTER_NORMALIZED_RESPONSE )
887                 prefilterNorm( *imgs0[i], *imgs[i], state->preFilterSize, state->preFilterCap, buf[i] );
888             else
889                 prefilterXSobel( *imgs0[i], *imgs[i], state->preFilterCap );
890         }
891     }
892 
893     const Mat* imgs0[2];
894     Mat* imgs[2];
895     uchar* buf[2];
896     StereoBMParams* state;
897 };
898 
ocl_stereobm(InputArray _left,InputArray _right,OutputArray _disp,StereoBMParams * state)899 static bool ocl_stereobm( InputArray _left, InputArray _right,
900                        OutputArray _disp, StereoBMParams* state)
901 {
902     int ndisp = state->numDisparities;
903     int mindisp = state->minDisparity;
904     int wsz = state->SADWindowSize;
905     int wsz2 = wsz/2;
906 
907     ocl::Device devDef = ocl::Device::getDefault();
908     int sizeX = devDef.isIntel() ? 32 : std::max(11, 27 - devDef.maxComputeUnits()),
909         sizeY = sizeX - 1,
910         N = ndisp * 2;
911 
912     cv::String opt = cv::format("-D DEFINE_KERNEL_STEREOBM -D MIN_DISP=%d -D NUM_DISP=%d"
913                                 " -D BLOCK_SIZE_X=%d -D BLOCK_SIZE_Y=%d -D WSZ=%d",
914                                 mindisp, ndisp,
915                                 sizeX, sizeY, wsz);
916     ocl::Kernel k("stereoBM", ocl::calib3d::stereobm_oclsrc, opt);
917     if(k.empty())
918         return false;
919 
920     UMat left = _left.getUMat(), right = _right.getUMat();
921     int cols = left.cols, rows = left.rows;
922 
923     _disp.create(_left.size(), CV_16S);
924     _disp.setTo((mindisp - 1) << 4);
925     Rect roi = Rect(Point(wsz2 + mindisp + ndisp - 1, wsz2), Point(cols-wsz2-mindisp, rows-wsz2) );
926     UMat disp = (_disp.getUMat())(roi);
927 
928     int globalX = (disp.cols + sizeX - 1) / sizeX,
929         globalY = (disp.rows + sizeY - 1) / sizeY;
930     size_t globalThreads[3] = {N, globalX, globalY};
931     size_t localThreads[3]  = {N, 1, 1};
932 
933     int idx = 0;
934     idx = k.set(idx, ocl::KernelArg::PtrReadOnly(left));
935     idx = k.set(idx, ocl::KernelArg::PtrReadOnly(right));
936     idx = k.set(idx, ocl::KernelArg::WriteOnlyNoSize(disp));
937     idx = k.set(idx, rows);
938     idx = k.set(idx, cols);
939     idx = k.set(idx, state->textureThreshold);
940     idx = k.set(idx, state->uniquenessRatio);
941     return k.run(3, globalThreads, localThreads, false);
942 }
943 
944 struct FindStereoCorrespInvoker : public ParallelLoopBody
945 {
FindStereoCorrespInvokercv::FindStereoCorrespInvoker946     FindStereoCorrespInvoker( const Mat& _left, const Mat& _right,
947                              Mat& _disp, StereoBMParams* _state,
948                              int _nstripes, size_t _stripeBufSize,
949                              bool _useShorts, Rect _validDisparityRect,
950                              Mat& _slidingSumBuf, Mat& _cost )
951     {
952         left = &_left; right = &_right;
953         disp = &_disp; state = _state;
954         nstripes = _nstripes; stripeBufSize = _stripeBufSize;
955         useShorts = _useShorts;
956         validDisparityRect = _validDisparityRect;
957         slidingSumBuf = &_slidingSumBuf;
958         cost = &_cost;
959     }
960 
operator ()cv::FindStereoCorrespInvoker961     void operator()( const Range& range ) const
962     {
963         int cols = left->cols, rows = left->rows;
964         int _row0 = std::min(cvRound(range.start * rows / nstripes), rows);
965         int _row1 = std::min(cvRound(range.end * rows / nstripes), rows);
966         uchar *ptr = slidingSumBuf->ptr() + range.start * stripeBufSize;
967         int FILTERED = (state->minDisparity - 1)*16;
968 
969         Rect roi = validDisparityRect & Rect(0, _row0, cols, _row1 - _row0);
970         if( roi.height == 0 )
971             return;
972         int row0 = roi.y;
973         int row1 = roi.y + roi.height;
974 
975         Mat part;
976         if( row0 > _row0 )
977         {
978             part = disp->rowRange(_row0, row0);
979             part = Scalar::all(FILTERED);
980         }
981         if( _row1 > row1 )
982         {
983             part = disp->rowRange(row1, _row1);
984             part = Scalar::all(FILTERED);
985         }
986 
987         Mat left_i = left->rowRange(row0, row1);
988         Mat right_i = right->rowRange(row0, row1);
989         Mat disp_i = disp->rowRange(row0, row1);
990         Mat cost_i = state->disp12MaxDiff >= 0 ? cost->rowRange(row0, row1) : Mat();
991 
992 #if CV_SSE2
993         if( useShorts )
994             findStereoCorrespondenceBM_SSE2( left_i, right_i, disp_i, cost_i, *state, ptr, row0, rows - row1 );
995         else
996 #endif
997             findStereoCorrespondenceBM( left_i, right_i, disp_i, cost_i, *state, ptr, row0, rows - row1 );
998 
999         if( state->disp12MaxDiff >= 0 )
1000             validateDisparity( disp_i, cost_i, state->minDisparity, state->numDisparities, state->disp12MaxDiff );
1001 
1002         if( roi.x > 0 )
1003         {
1004             part = disp_i.colRange(0, roi.x);
1005             part = Scalar::all(FILTERED);
1006         }
1007         if( roi.x + roi.width < cols )
1008         {
1009             part = disp_i.colRange(roi.x + roi.width, cols);
1010             part = Scalar::all(FILTERED);
1011         }
1012     }
1013 
1014 protected:
1015     const Mat *left, *right;
1016     Mat* disp, *slidingSumBuf, *cost;
1017     StereoBMParams *state;
1018 
1019     int nstripes;
1020     size_t stripeBufSize;
1021     bool useShorts;
1022     Rect validDisparityRect;
1023 };
1024 
1025 class StereoBMImpl : public StereoBM
1026 {
1027 public:
StereoBMImpl()1028     StereoBMImpl()
1029     {
1030         params = StereoBMParams();
1031     }
1032 
StereoBMImpl(int _numDisparities,int _SADWindowSize)1033     StereoBMImpl( int _numDisparities, int _SADWindowSize )
1034     {
1035         params = StereoBMParams(_numDisparities, _SADWindowSize);
1036     }
1037 
compute(InputArray leftarr,InputArray rightarr,OutputArray disparr)1038     void compute( InputArray leftarr, InputArray rightarr, OutputArray disparr )
1039     {
1040         int dtype = disparr.fixedType() ? disparr.type() : params.dispType;
1041         Size leftsize = leftarr.size();
1042 
1043         if (leftarr.size() != rightarr.size())
1044             CV_Error( Error::StsUnmatchedSizes, "All the images must have the same size" );
1045 
1046         if (leftarr.type() != CV_8UC1 || rightarr.type() != CV_8UC1)
1047             CV_Error( Error::StsUnsupportedFormat, "Both input images must have CV_8UC1" );
1048 
1049         if (dtype != CV_16SC1 && dtype != CV_32FC1)
1050             CV_Error( Error::StsUnsupportedFormat, "Disparity image must have CV_16SC1 or CV_32FC1 format" );
1051 
1052         if( params.preFilterType != PREFILTER_NORMALIZED_RESPONSE &&
1053             params.preFilterType != PREFILTER_XSOBEL )
1054             CV_Error( Error::StsOutOfRange, "preFilterType must be = CV_STEREO_BM_NORMALIZED_RESPONSE" );
1055 
1056         if( params.preFilterSize < 5 || params.preFilterSize > 255 || params.preFilterSize % 2 == 0 )
1057             CV_Error( Error::StsOutOfRange, "preFilterSize must be odd and be within 5..255" );
1058 
1059         if( params.preFilterCap < 1 || params.preFilterCap > 63 )
1060             CV_Error( Error::StsOutOfRange, "preFilterCap must be within 1..63" );
1061 
1062         if( params.SADWindowSize < 5 || params.SADWindowSize > 255 || params.SADWindowSize % 2 == 0 ||
1063             params.SADWindowSize >= std::min(leftsize.width, leftsize.height) )
1064             CV_Error( Error::StsOutOfRange, "SADWindowSize must be odd, be within 5..255 and be not larger than image width or height" );
1065 
1066         if( params.numDisparities <= 0 || params.numDisparities % 16 != 0 )
1067             CV_Error( Error::StsOutOfRange, "numDisparities must be positive and divisble by 16" );
1068 
1069         if( params.textureThreshold < 0 )
1070             CV_Error( Error::StsOutOfRange, "texture threshold must be non-negative" );
1071 
1072         if( params.uniquenessRatio < 0 )
1073             CV_Error( Error::StsOutOfRange, "uniqueness ratio must be non-negative" );
1074 
1075         int FILTERED = (params.minDisparity - 1) << DISPARITY_SHIFT;
1076 
1077         if(ocl::useOpenCL() && disparr.isUMat() && params.textureThreshold == 0)
1078         {
1079             UMat left, right;
1080             if(ocl_prefiltering(leftarr, rightarr, left, right, &params))
1081             {
1082                 if(ocl_stereobm(left, right, disparr, &params))
1083                 {
1084                     if( params.speckleRange >= 0 && params.speckleWindowSize > 0 )
1085                         filterSpeckles(disparr.getMat(), FILTERED, params.speckleWindowSize, params.speckleRange, slidingSumBuf);
1086                     if (dtype == CV_32F)
1087                         disparr.getUMat().convertTo(disparr, CV_32FC1, 1./(1 << DISPARITY_SHIFT), 0);
1088                     CV_IMPL_ADD(CV_IMPL_OCL);
1089                     return;
1090                 }
1091             }
1092         }
1093 
1094         Mat left0 = leftarr.getMat(), right0 = rightarr.getMat();
1095         disparr.create(left0.size(), dtype);
1096         Mat disp0 = disparr.getMat();
1097 
1098         preFilteredImg0.create( left0.size(), CV_8U );
1099         preFilteredImg1.create( left0.size(), CV_8U );
1100         cost.create( left0.size(), CV_16S );
1101 
1102         Mat left = preFilteredImg0, right = preFilteredImg1;
1103 
1104         int mindisp = params.minDisparity;
1105         int ndisp = params.numDisparities;
1106 
1107         int width = left0.cols;
1108         int height = left0.rows;
1109         int lofs = std::max(ndisp - 1 + mindisp, 0);
1110         int rofs = -std::min(ndisp - 1 + mindisp, 0);
1111         int width1 = width - rofs - ndisp + 1;
1112 
1113         if( lofs >= width || rofs >= width || width1 < 1 )
1114         {
1115             disp0 = Scalar::all( FILTERED * ( disp0.type() < CV_32F ? 1 : 1./(1 << DISPARITY_SHIFT) ) );
1116             return;
1117         }
1118 
1119         Mat disp = disp0;
1120         if( dtype == CV_32F )
1121         {
1122             dispbuf.create(disp0.size(), CV_16S);
1123             disp = dispbuf;
1124         }
1125 
1126         int wsz = params.SADWindowSize;
1127         int bufSize0 = (int)((ndisp + 2)*sizeof(int));
1128         bufSize0 += (int)((height+wsz+2)*ndisp*sizeof(int));
1129         bufSize0 += (int)((height + wsz + 2)*sizeof(int));
1130         bufSize0 += (int)((height+wsz+2)*ndisp*(wsz+2)*sizeof(uchar) + 256);
1131 
1132         int bufSize1 = (int)((width + params.preFilterSize + 2) * sizeof(int) + 256);
1133         int bufSize2 = 0;
1134         if( params.speckleRange >= 0 && params.speckleWindowSize > 0 )
1135             bufSize2 = width*height*(sizeof(Point_<short>) + sizeof(int) + sizeof(uchar));
1136 
1137 #if CV_SSE2
1138         bool useShorts = params.preFilterCap <= 31 && params.SADWindowSize <= 21 && checkHardwareSupport(CV_CPU_SSE2);
1139 #else
1140         const bool useShorts = false;
1141 #endif
1142 
1143         const double SAD_overhead_coeff = 10.0;
1144         double N0 = 8000000 / (useShorts ? 1 : 4);  // approx tbb's min number instructions reasonable for one thread
1145         double maxStripeSize = std::min(std::max(N0 / (width * ndisp), (wsz-1) * SAD_overhead_coeff), (double)height);
1146         int nstripes = cvCeil(height / maxStripeSize);
1147         int bufSize = std::max(bufSize0 * nstripes, std::max(bufSize1 * 2, bufSize2));
1148 
1149         if( slidingSumBuf.cols < bufSize )
1150             slidingSumBuf.create( 1, bufSize, CV_8U );
1151 
1152         uchar *_buf = slidingSumBuf.ptr();
1153 
1154         parallel_for_(Range(0, 2), PrefilterInvoker(left0, right0, left, right, _buf, _buf + bufSize1, &params), 1);
1155 
1156         Rect validDisparityRect(0, 0, width, height), R1 = params.roi1, R2 = params.roi2;
1157         validDisparityRect = getValidDisparityROI(R1.area() > 0 ? Rect(0, 0, width, height) : validDisparityRect,
1158                                                   R2.area() > 0 ? Rect(0, 0, width, height) : validDisparityRect,
1159                                                   params.minDisparity, params.numDisparities,
1160                                                   params.SADWindowSize);
1161 
1162         parallel_for_(Range(0, nstripes),
1163                       FindStereoCorrespInvoker(left, right, disp, &params, nstripes,
1164                                                bufSize0, useShorts, validDisparityRect,
1165                                                slidingSumBuf, cost));
1166 
1167         if( params.speckleRange >= 0 && params.speckleWindowSize > 0 )
1168             filterSpeckles(disp, FILTERED, params.speckleWindowSize, params.speckleRange, slidingSumBuf);
1169 
1170         if (disp0.data != disp.data)
1171             disp.convertTo(disp0, disp0.type(), 1./(1 << DISPARITY_SHIFT), 0);
1172     }
1173 
getMinDisparity() const1174     int getMinDisparity() const { return params.minDisparity; }
setMinDisparity(int minDisparity)1175     void setMinDisparity(int minDisparity) { params.minDisparity = minDisparity; }
1176 
getNumDisparities() const1177     int getNumDisparities() const { return params.numDisparities; }
setNumDisparities(int numDisparities)1178     void setNumDisparities(int numDisparities) { params.numDisparities = numDisparities; }
1179 
getBlockSize() const1180     int getBlockSize() const { return params.SADWindowSize; }
setBlockSize(int blockSize)1181     void setBlockSize(int blockSize) { params.SADWindowSize = blockSize; }
1182 
getSpeckleWindowSize() const1183     int getSpeckleWindowSize() const { return params.speckleWindowSize; }
setSpeckleWindowSize(int speckleWindowSize)1184     void setSpeckleWindowSize(int speckleWindowSize) { params.speckleWindowSize = speckleWindowSize; }
1185 
getSpeckleRange() const1186     int getSpeckleRange() const { return params.speckleRange; }
setSpeckleRange(int speckleRange)1187     void setSpeckleRange(int speckleRange) { params.speckleRange = speckleRange; }
1188 
getDisp12MaxDiff() const1189     int getDisp12MaxDiff() const { return params.disp12MaxDiff; }
setDisp12MaxDiff(int disp12MaxDiff)1190     void setDisp12MaxDiff(int disp12MaxDiff) { params.disp12MaxDiff = disp12MaxDiff; }
1191 
getPreFilterType() const1192     int getPreFilterType() const { return params.preFilterType; }
setPreFilterType(int preFilterType)1193     void setPreFilterType(int preFilterType) { params.preFilterType = preFilterType; }
1194 
getPreFilterSize() const1195     int getPreFilterSize() const { return params.preFilterSize; }
setPreFilterSize(int preFilterSize)1196     void setPreFilterSize(int preFilterSize) { params.preFilterSize = preFilterSize; }
1197 
getPreFilterCap() const1198     int getPreFilterCap() const { return params.preFilterCap; }
setPreFilterCap(int preFilterCap)1199     void setPreFilterCap(int preFilterCap) { params.preFilterCap = preFilterCap; }
1200 
getTextureThreshold() const1201     int getTextureThreshold() const { return params.textureThreshold; }
setTextureThreshold(int textureThreshold)1202     void setTextureThreshold(int textureThreshold) { params.textureThreshold = textureThreshold; }
1203 
getUniquenessRatio() const1204     int getUniquenessRatio() const { return params.uniquenessRatio; }
setUniquenessRatio(int uniquenessRatio)1205     void setUniquenessRatio(int uniquenessRatio) { params.uniquenessRatio = uniquenessRatio; }
1206 
getSmallerBlockSize() const1207     int getSmallerBlockSize() const { return 0; }
setSmallerBlockSize(int)1208     void setSmallerBlockSize(int) {}
1209 
getROI1() const1210     Rect getROI1() const { return params.roi1; }
setROI1(Rect roi1)1211     void setROI1(Rect roi1) { params.roi1 = roi1; }
1212 
getROI2() const1213     Rect getROI2() const { return params.roi2; }
setROI2(Rect roi2)1214     void setROI2(Rect roi2) { params.roi2 = roi2; }
1215 
write(FileStorage & fs) const1216     void write(FileStorage& fs) const
1217     {
1218         fs << "name" << name_
1219         << "minDisparity" << params.minDisparity
1220         << "numDisparities" << params.numDisparities
1221         << "blockSize" << params.SADWindowSize
1222         << "speckleWindowSize" << params.speckleWindowSize
1223         << "speckleRange" << params.speckleRange
1224         << "disp12MaxDiff" << params.disp12MaxDiff
1225         << "preFilterType" << params.preFilterType
1226         << "preFilterSize" << params.preFilterSize
1227         << "preFilterCap" << params.preFilterCap
1228         << "textureThreshold" << params.textureThreshold
1229         << "uniquenessRatio" << params.uniquenessRatio;
1230     }
1231 
read(const FileNode & fn)1232     void read(const FileNode& fn)
1233     {
1234         FileNode n = fn["name"];
1235         CV_Assert( n.isString() && String(n) == name_ );
1236         params.minDisparity = (int)fn["minDisparity"];
1237         params.numDisparities = (int)fn["numDisparities"];
1238         params.SADWindowSize = (int)fn["blockSize"];
1239         params.speckleWindowSize = (int)fn["speckleWindowSize"];
1240         params.speckleRange = (int)fn["speckleRange"];
1241         params.disp12MaxDiff = (int)fn["disp12MaxDiff"];
1242         params.preFilterType = (int)fn["preFilterType"];
1243         params.preFilterSize = (int)fn["preFilterSize"];
1244         params.preFilterCap = (int)fn["preFilterCap"];
1245         params.textureThreshold = (int)fn["textureThreshold"];
1246         params.uniquenessRatio = (int)fn["uniquenessRatio"];
1247         params.roi1 = params.roi2 = Rect();
1248     }
1249 
1250     StereoBMParams params;
1251     Mat preFilteredImg0, preFilteredImg1, cost, dispbuf;
1252     Mat slidingSumBuf;
1253 
1254     static const char* name_;
1255 };
1256 
1257 const char* StereoBMImpl::name_ = "StereoMatcher.BM";
1258 
create(int _numDisparities,int _SADWindowSize)1259 Ptr<StereoBM> StereoBM::create(int _numDisparities, int _SADWindowSize)
1260 {
1261     return makePtr<StereoBMImpl>(_numDisparities, _SADWindowSize);
1262 }
1263 
1264 }
1265 
1266 /* End of file. */
1267