CoCalc -- convert

GitHub Repository: Tetragramm/opencv
Path: blob/master/modules/core/src/convert_scale.cpp
¹⁶³³⁷ views
1
// This file is part of OpenCV project.
2
// It is subject to the license terms in the LICENSE file found in the top-level directory
3
// of this distribution and at http://opencv.org/license.html
4

5

6
#include "precomp.hpp"
7
#include "opencl_kernels_core.hpp"
8
#include "convert.hpp"
9

10
/****************************************************************************************\
11
*                                convertScale[Abs]                                       *
12
\****************************************************************************************/
13

14
namespace cv
15
{
16

17
template<typename _Ts, typename _Td> inline void
18
cvtabs_32f( const _Ts* src, size_t sstep, _Td* dst, size_t dstep,
19
            Size size, float a, float b )
20
{
21
#if CV_SIMD
22
    v_float32 va = vx_setall_f32(a), vb = vx_setall_f32(b);
23
    const int VECSZ = v_float32::nlanes*2;
24
#endif
25
    sstep /= sizeof(src[0]);
26
    dstep /= sizeof(dst[0]);
27

28
    for( int i = 0; i < size.height; i++, src += sstep, dst += dstep )
29
    {
30
        int j = 0;
31
#if CV_SIMD
32
        for( ; j < size.width; j += VECSZ )
33
        {
34
            if( j > size.width - VECSZ )
35
            {
36
                if( j == 0 || src == (_Ts*)dst )
37
                    break;
38
                j = size.width - VECSZ;
39
            }
40
            v_float32 v0, v1;
41
            vx_load_pair_as(src + j, v0, v1);
42
            v0 = v_fma(v0, va, vb);
43
            v1 = v_fma(v1, va, vb);
44
            v_store_pair_as(dst + j, v_abs(v0), v_abs(v1));
45
        }
46
#endif
47
        for( ; j < size.width; j++ )
48
            dst[j] = saturate_cast<_Td>(std::abs(src[j]*a + b));
49
    }
50
}
51

52
// variant for convrsions 16f <-> ... w/o unrolling
53
template<typename _Ts, typename _Td> inline void
54
cvtabs1_32f( const _Ts* src, size_t sstep, _Td* dst, size_t dstep,
55
             Size size, float a, float b )
56
{
57
#if CV_SIMD
58
    v_float32 va = vx_setall_f32(a), vb = vx_setall_f32(b);
59
    const int VECSZ = v_float32::nlanes*2;
60
#endif
61
    sstep /= sizeof(src[0]);
62
    dstep /= sizeof(dst[0]);
63

64
    for( int i = 0; i < size.height; i++, src += sstep, dst += dstep )
65
    {
66
        int j = 0;
67
#if CV_SIMD
68
        for( ; j < size.width; j += VECSZ )
69
        {
70
            if( j > size.width - VECSZ )
71
            {
72
                if( j == 0 || src == (_Ts*)dst )
73
                    break;
74
                j = size.width - VECSZ;
75
            }
76
            v_float32 v0;
77
            vx_load_as(src + j, v0);
78
            v0 = v_fma(v0, va, vb);
79
            v_store_as(dst + j, v_abs(v0));
80
        }
81
#endif
82
        for( ; j < size.width; j++ )
83
            dst[j] = saturate_cast<_Td>(src[j]*a + b);
84
    }
85
}
86

87
template<typename _Ts, typename _Td> inline void
88
cvt_32f( const _Ts* src, size_t sstep, _Td* dst, size_t dstep,
89
         Size size, float a, float b )
90
{
91
#if CV_SIMD
92
    v_float32 va = vx_setall_f32(a), vb = vx_setall_f32(b);
93
    const int VECSZ = v_float32::nlanes*2;
94
#endif
95
    sstep /= sizeof(src[0]);
96
    dstep /= sizeof(dst[0]);
97

98
    for( int i = 0; i < size.height; i++, src += sstep, dst += dstep )
99
    {
100
        int j = 0;
101
#if CV_SIMD
102
        for( ; j < size.width; j += VECSZ )
103
        {
104
            if( j > size.width - VECSZ )
105
            {
106
                if( j == 0 || src == (_Ts*)dst )
107
                    break;
108
                j = size.width - VECSZ;
109
            }
110
            v_float32 v0, v1;
111
            vx_load_pair_as(src + j, v0, v1);
112
            v0 = v_fma(v0, va, vb);
113
            v1 = v_fma(v1, va, vb);
114
            v_store_pair_as(dst + j, v0, v1);
115
        }
116
#endif
117
        for( ; j < size.width; j++ )
118
            dst[j] = saturate_cast<_Td>(src[j]*a + b);
119
    }
120
}
121

122
// variant for convrsions 16f <-> ... w/o unrolling
123
template<typename _Ts, typename _Td> inline void
124
cvt1_32f( const _Ts* src, size_t sstep, _Td* dst, size_t dstep,
125
          Size size, float a, float b )
126
{
127
#if CV_SIMD
128
    v_float32 va = vx_setall_f32(a), vb = vx_setall_f32(b);
129
    const int VECSZ = v_float32::nlanes;
130
#endif
131
    sstep /= sizeof(src[0]);
132
    dstep /= sizeof(dst[0]);
133

134
    for( int i = 0; i < size.height; i++, src += sstep, dst += dstep )
135
    {
136
        int j = 0;
137
#if CV_SIMD
138
        for( ; j < size.width; j += VECSZ )
139
        {
140
            if( j > size.width - VECSZ )
141
            {
142
                if( j == 0 || src == (_Ts*)dst )
143
                    break;
144
                j = size.width - VECSZ;
145
            }
146
            v_float32 v0;
147
            vx_load_as(src + j, v0);
148
            v0 = v_fma(v0, va, vb);
149
            v_store_as(dst + j, v0);
150
        }
151
#endif
152
        for( ; j < size.width; j++ )
153
            dst[j] = saturate_cast<_Td>(src[j]*a + b);
154
    }
155
}
156

157

158
template<typename _Ts, typename _Td> inline void
159
cvt_64f( const _Ts* src, size_t sstep, _Td* dst, size_t dstep,
160
         Size size, double a, double b )
161
{
162
#if CV_SIMD_64F
163
    v_float64 va = vx_setall_f64(a), vb = vx_setall_f64(b);
164
    const int VECSZ = v_float64::nlanes*2;
165
#endif
166
    sstep /= sizeof(src[0]);
167
    dstep /= sizeof(dst[0]);
168

169
    for( int i = 0; i < size.height; i++, src += sstep, dst += dstep )
170
    {
171
        int j = 0;
172
#if CV_SIMD_64F
173
        for( ; j < size.width; j += VECSZ )
174
        {
175
            if( j > size.width - VECSZ )
176
            {
177
                if( j == 0 || src == (_Ts*)dst )
178
                    break;
179
                j = size.width - VECSZ;
180
            }
181
            v_float64 v0, v1;
182
            vx_load_pair_as(src + j, v0, v1);
183
            v0 = v_fma(v0, va, vb);
184
            v1 = v_fma(v1, va, vb);
185
            v_store_pair_as(dst + j, v0, v1);
186
        }
187
#endif
188
        for( ; j < size.width; j++ )
189
            dst[j] = saturate_cast<_Td>(src[j]*a + b);
190
    }
191
}
192

193
//==================================================================================================
194

195
#define DEF_CVT_SCALE_ABS_FUNC(suffix, cvt, stype, dtype, wtype) \
196
static void cvtScaleAbs##suffix( const stype* src, size_t sstep, const uchar*, size_t, \
197
                                 dtype* dst, size_t dstep, Size size, double* scale) \
198
{ \
199
    cvt(src, sstep, dst, dstep, size, (wtype)scale[0], (wtype)scale[1]); \
200
}
201

202

203
#define DEF_CVT_SCALE_FUNC(suffix, cvt, stype, dtype, wtype) \
204
static void cvtScale##suffix( const stype* src, size_t sstep, const uchar*, size_t, \
205
                              dtype* dst, size_t dstep, Size size, double* scale) \
206
{ \
207
    cvt(src, sstep, dst, dstep, size, (wtype)scale[0], (wtype)scale[1]); \
208
}
209

210
DEF_CVT_SCALE_ABS_FUNC(8u,    cvtabs_32f, uchar,  uchar, float)
211
DEF_CVT_SCALE_ABS_FUNC(8s8u,  cvtabs_32f, schar,  uchar, float)
212
DEF_CVT_SCALE_ABS_FUNC(16u8u, cvtabs_32f, ushort, uchar, float)
213
DEF_CVT_SCALE_ABS_FUNC(16s8u, cvtabs_32f, short,  uchar, float)
214
DEF_CVT_SCALE_ABS_FUNC(32s8u, cvtabs_32f, int,    uchar, float)
215
DEF_CVT_SCALE_ABS_FUNC(32f8u, cvtabs_32f, float,  uchar, float)
216
DEF_CVT_SCALE_ABS_FUNC(64f8u, cvtabs_32f, double, uchar, float)
217

218
DEF_CVT_SCALE_FUNC(8u,     cvt_32f, uchar,  uchar, float)
219
DEF_CVT_SCALE_FUNC(8s8u,   cvt_32f, schar,  uchar, float)
220
DEF_CVT_SCALE_FUNC(16u8u,  cvt_32f, ushort, uchar, float)
221
DEF_CVT_SCALE_FUNC(16s8u,  cvt_32f, short,  uchar, float)
222
DEF_CVT_SCALE_FUNC(32s8u,  cvt_32f, int,    uchar, float)
223
DEF_CVT_SCALE_FUNC(32f8u,  cvt_32f, float,  uchar, float)
224
DEF_CVT_SCALE_FUNC(64f8u,  cvt_32f, double, uchar, float)
225
DEF_CVT_SCALE_FUNC(16f8u,  cvt_32f, float16_t, uchar, float)
226

227
DEF_CVT_SCALE_FUNC(8u8s,   cvt_32f, uchar,  schar, float)
228
DEF_CVT_SCALE_FUNC(8s,     cvt_32f, schar,  schar, float)
229
DEF_CVT_SCALE_FUNC(16u8s,  cvt_32f, ushort, schar, float)
230
DEF_CVT_SCALE_FUNC(16s8s,  cvt_32f, short,  schar, float)
231
DEF_CVT_SCALE_FUNC(32s8s,  cvt_32f, int,    schar, float)
232
DEF_CVT_SCALE_FUNC(32f8s,  cvt_32f, float,  schar, float)
233
DEF_CVT_SCALE_FUNC(64f8s,  cvt_32f, double, schar, float)
234
DEF_CVT_SCALE_FUNC(16f8s,  cvt_32f, float16_t, schar, float)
235

236
DEF_CVT_SCALE_FUNC(8u16u,  cvt_32f, uchar,  ushort, float)
237
DEF_CVT_SCALE_FUNC(8s16u,  cvt_32f, schar,  ushort, float)
238
DEF_CVT_SCALE_FUNC(16u,    cvt_32f, ushort, ushort, float)
239
DEF_CVT_SCALE_FUNC(16s16u, cvt_32f, short,  ushort, float)
240
DEF_CVT_SCALE_FUNC(32s16u, cvt_32f, int,    ushort, float)
241
DEF_CVT_SCALE_FUNC(32f16u, cvt_32f, float,  ushort, float)
242
DEF_CVT_SCALE_FUNC(64f16u, cvt_32f, double, ushort, float)
243
DEF_CVT_SCALE_FUNC(16f16u, cvt1_32f, float16_t, ushort, float)
244

245
DEF_CVT_SCALE_FUNC(8u16s,  cvt_32f, uchar,  short, float)
246
DEF_CVT_SCALE_FUNC(8s16s,  cvt_32f, schar,  short, float)
247
DEF_CVT_SCALE_FUNC(16u16s, cvt_32f, ushort, short, float)
248
DEF_CVT_SCALE_FUNC(16s,    cvt_32f, short,  short, float)
249
DEF_CVT_SCALE_FUNC(32s16s, cvt_32f, int,    short, float)
250
DEF_CVT_SCALE_FUNC(32f16s, cvt_32f, float,  short, float)
251
DEF_CVT_SCALE_FUNC(64f16s, cvt_32f, double, short, float)
252
DEF_CVT_SCALE_FUNC(16f16s, cvt1_32f, float16_t, short, float)
253

254
DEF_CVT_SCALE_FUNC(8u32s,  cvt_32f, uchar,  int, float)
255
DEF_CVT_SCALE_FUNC(8s32s,  cvt_32f, schar,  int, float)
256
DEF_CVT_SCALE_FUNC(16u32s, cvt_32f, ushort, int, float)
257
DEF_CVT_SCALE_FUNC(16s32s, cvt_32f, short,  int, float)
258
DEF_CVT_SCALE_FUNC(32s,    cvt_64f, int,    int, double)
259
DEF_CVT_SCALE_FUNC(32f32s, cvt_32f, float,  int, float)
260
DEF_CVT_SCALE_FUNC(64f32s, cvt_64f, double, int, double)
261
DEF_CVT_SCALE_FUNC(16f32s, cvt1_32f, float16_t, int, float)
262

263
DEF_CVT_SCALE_FUNC(8u32f,  cvt_32f, uchar,  float, float)
264
DEF_CVT_SCALE_FUNC(8s32f,  cvt_32f, schar,  float, float)
265
DEF_CVT_SCALE_FUNC(16u32f, cvt_32f, ushort, float, float)
266
DEF_CVT_SCALE_FUNC(16s32f, cvt_32f, short,  float, float)
267
DEF_CVT_SCALE_FUNC(32s32f, cvt_32f, int,    float, float)
268
DEF_CVT_SCALE_FUNC(32f,    cvt_32f, float,  float, float)
269
DEF_CVT_SCALE_FUNC(64f32f, cvt_64f, double, float, double)
270
DEF_CVT_SCALE_FUNC(16f32f, cvt1_32f, float16_t, float, float)
271

272
DEF_CVT_SCALE_FUNC(8u64f,  cvt_64f, uchar,  double, double)
273
DEF_CVT_SCALE_FUNC(8s64f,  cvt_64f, schar,  double, double)
274
DEF_CVT_SCALE_FUNC(16u64f, cvt_64f, ushort, double, double)
275
DEF_CVT_SCALE_FUNC(16s64f, cvt_64f, short,  double, double)
276
DEF_CVT_SCALE_FUNC(32s64f, cvt_64f, int,    double, double)
277
DEF_CVT_SCALE_FUNC(32f64f, cvt_64f, float,  double, double)
278
DEF_CVT_SCALE_FUNC(64f,    cvt_64f, double, double, double)
279
DEF_CVT_SCALE_FUNC(16f64f, cvt_64f, float16_t, double, double)
280

281
DEF_CVT_SCALE_FUNC(8u16f,  cvt1_32f, uchar,  float16_t, float)
282
DEF_CVT_SCALE_FUNC(8s16f,  cvt1_32f, schar,  float16_t, float)
283
DEF_CVT_SCALE_FUNC(16u16f, cvt1_32f, ushort, float16_t, float)
284
DEF_CVT_SCALE_FUNC(16s16f, cvt1_32f, short,  float16_t, float)
285
DEF_CVT_SCALE_FUNC(32s16f, cvt1_32f, int,    float16_t, float)
286
DEF_CVT_SCALE_FUNC(32f16f, cvt1_32f, float,  float16_t, float)
287
DEF_CVT_SCALE_FUNC(64f16f, cvt_64f,  double, float16_t, double)
288
DEF_CVT_SCALE_FUNC(16f,    cvt1_32f, float16_t, float16_t, float)
289

290
static BinaryFunc getCvtScaleAbsFunc(int depth)
291
{
292
    static BinaryFunc cvtScaleAbsTab[] =
293
    {
294
        (BinaryFunc)cvtScaleAbs8u, (BinaryFunc)cvtScaleAbs8s8u, (BinaryFunc)cvtScaleAbs16u8u,
295
        (BinaryFunc)cvtScaleAbs16s8u, (BinaryFunc)cvtScaleAbs32s8u, (BinaryFunc)cvtScaleAbs32f8u,
296
        (BinaryFunc)cvtScaleAbs64f8u, 0
297
    };
298

299
    return cvtScaleAbsTab[depth];
300
}
301

302
BinaryFunc getConvertScaleFunc(int sdepth, int ddepth)
303
{
304
    static BinaryFunc cvtScaleTab[][8] =
305
    {
306
        {
307
            (BinaryFunc)GET_OPTIMIZED(cvtScale8u), (BinaryFunc)GET_OPTIMIZED(cvtScale8s8u), (BinaryFunc)GET_OPTIMIZED(cvtScale16u8u),
308
            (BinaryFunc)GET_OPTIMIZED(cvtScale16s8u), (BinaryFunc)GET_OPTIMIZED(cvtScale32s8u), (BinaryFunc)GET_OPTIMIZED(cvtScale32f8u),
309
            (BinaryFunc)cvtScale64f8u, (BinaryFunc)cvtScale16f8u
310
        },
311
        {
312
            (BinaryFunc)GET_OPTIMIZED(cvtScale8u8s), (BinaryFunc)GET_OPTIMIZED(cvtScale8s), (BinaryFunc)GET_OPTIMIZED(cvtScale16u8s),
313
            (BinaryFunc)GET_OPTIMIZED(cvtScale16s8s), (BinaryFunc)GET_OPTIMIZED(cvtScale32s8s), (BinaryFunc)GET_OPTIMIZED(cvtScale32f8s),
314
            (BinaryFunc)cvtScale64f8s, (BinaryFunc)cvtScale16f8s
315
        },
316
        {
317
            (BinaryFunc)GET_OPTIMIZED(cvtScale8u16u), (BinaryFunc)GET_OPTIMIZED(cvtScale8s16u), (BinaryFunc)GET_OPTIMIZED(cvtScale16u),
318
            (BinaryFunc)GET_OPTIMIZED(cvtScale16s16u), (BinaryFunc)GET_OPTIMIZED(cvtScale32s16u), (BinaryFunc)GET_OPTIMIZED(cvtScale32f16u),
319
            (BinaryFunc)cvtScale64f16u, (BinaryFunc)cvtScale16f16u
320
        },
321
        {
322
            (BinaryFunc)GET_OPTIMIZED(cvtScale8u16s), (BinaryFunc)GET_OPTIMIZED(cvtScale8s16s), (BinaryFunc)GET_OPTIMIZED(cvtScale16u16s),
323
            (BinaryFunc)GET_OPTIMIZED(cvtScale16s), (BinaryFunc)GET_OPTIMIZED(cvtScale32s16s), (BinaryFunc)GET_OPTIMIZED(cvtScale32f16s),
324
            (BinaryFunc)cvtScale64f16s, (BinaryFunc)cvtScale16f16s
325
        },
326
        {
327
            (BinaryFunc)GET_OPTIMIZED(cvtScale8u32s), (BinaryFunc)GET_OPTIMIZED(cvtScale8s32s), (BinaryFunc)GET_OPTIMIZED(cvtScale16u32s),
328
            (BinaryFunc)GET_OPTIMIZED(cvtScale16s32s), (BinaryFunc)GET_OPTIMIZED(cvtScale32s), (BinaryFunc)GET_OPTIMIZED(cvtScale32f32s),
329
            (BinaryFunc)cvtScale64f32s, (BinaryFunc)cvtScale16f32s
330
        },
331
        {
332
            (BinaryFunc)GET_OPTIMIZED(cvtScale8u32f), (BinaryFunc)GET_OPTIMIZED(cvtScale8s32f), (BinaryFunc)GET_OPTIMIZED(cvtScale16u32f),
333
            (BinaryFunc)GET_OPTIMIZED(cvtScale16s32f), (BinaryFunc)GET_OPTIMIZED(cvtScale32s32f), (BinaryFunc)GET_OPTIMIZED(cvtScale32f),
334
            (BinaryFunc)cvtScale64f32f, (BinaryFunc)cvtScale16f32f
335
        },
336
        {
337
            (BinaryFunc)cvtScale8u64f, (BinaryFunc)cvtScale8s64f, (BinaryFunc)cvtScale16u64f,
338
            (BinaryFunc)cvtScale16s64f, (BinaryFunc)cvtScale32s64f, (BinaryFunc)cvtScale32f64f,
339
            (BinaryFunc)cvtScale64f, (BinaryFunc)cvtScale16f64f
340
        },
341
        {
342
            (BinaryFunc)cvtScale8u16f, (BinaryFunc)cvtScale8s16f, (BinaryFunc)cvtScale16u16f,
343
            (BinaryFunc)cvtScale16s16f, (BinaryFunc)cvtScale32s16f, (BinaryFunc)cvtScale32f16f,
344
            (BinaryFunc)cvtScale64f16f, (BinaryFunc)cvtScale16f
345
        },
346
    };
347

348
    return cvtScaleTab[CV_MAT_DEPTH(ddepth)][CV_MAT_DEPTH(sdepth)];
349
}
350

351
#ifdef HAVE_OPENCL
352

353
static bool ocl_convertScaleAbs( InputArray _src, OutputArray _dst, double alpha, double beta )
354
{
355
    const ocl::Device & d = ocl::Device::getDefault();
356

357
    int type = _src.type(), depth = CV_MAT_DEPTH(type), cn = CV_MAT_CN(type);
358
    bool doubleSupport = d.doubleFPConfig() > 0;
359
    if (!doubleSupport && depth == CV_64F)
360
        return false;
361

362
    _dst.create(_src.size(), CV_8UC(cn));
363
    int kercn = 1;
364
    if (d.isIntel())
365
    {
366
        static const int vectorWidths[] = {4, 4, 4, 4, 4, 4, 4, -1};
367
        kercn = ocl::checkOptimalVectorWidth( vectorWidths, _src, _dst,
368
                                              noArray(), noArray(), noArray(),
369
                                              noArray(), noArray(), noArray(),
370
                                              noArray(), ocl::OCL_VECTOR_MAX);
371
    }
372
    else
373
        kercn = ocl::predictOptimalVectorWidthMax(_src, _dst);
374

375
    int rowsPerWI = d.isIntel() ? 4 : 1;
376
    char cvt[2][50];
377
    int wdepth = std::max(depth, CV_32F);
378
    String build_opt = format("-D OP_CONVERT_SCALE_ABS -D UNARY_OP -D dstT=%s -D DEPTH_dst=%d -D srcT1=%s"
379
                         " -D workT=%s -D wdepth=%d -D convertToWT1=%s -D convertToDT=%s"
380
                         " -D workT1=%s -D rowsPerWI=%d%s",
381
                         ocl::typeToStr(CV_8UC(kercn)), CV_8U,
382
                         ocl::typeToStr(CV_MAKE_TYPE(depth, kercn)),
383
                         ocl::typeToStr(CV_MAKE_TYPE(wdepth, kercn)), wdepth,
384
                         ocl::convertTypeStr(depth, wdepth, kercn, cvt[0]),
385
                         ocl::convertTypeStr(wdepth, CV_8U, kercn, cvt[1]),
386
                         ocl::typeToStr(wdepth), rowsPerWI,
387
                         doubleSupport ? " -D DOUBLE_SUPPORT" : "");
388
    ocl::Kernel k("KF", ocl::core::arithm_oclsrc, build_opt);
389
    if (k.empty())
390
        return false;
391

392
    UMat src = _src.getUMat();
393
    UMat dst = _dst.getUMat();
394

395
    ocl::KernelArg srcarg = ocl::KernelArg::ReadOnlyNoSize(src),
396
            dstarg = ocl::KernelArg::WriteOnly(dst, cn, kercn);
397

398
    if (wdepth == CV_32F)
399
        k.args(srcarg, dstarg, (float)alpha, (float)beta);
400
    else if (wdepth == CV_64F)
401
        k.args(srcarg, dstarg, alpha, beta);
402

403
    size_t globalsize[2] = { (size_t)src.cols * cn / kercn, ((size_t)src.rows + rowsPerWI - 1) / rowsPerWI };
404
    return k.run(2, globalsize, NULL, false);
405
}
406

407
#endif
408

409
} //cv::
410

411

412
void cv::convertScaleAbs( InputArray _src, OutputArray _dst, double alpha, double beta )
413
{
414
    CV_INSTRUMENT_REGION();
415

416
    CV_OCL_RUN(_src.dims() <= 2 && _dst.isUMat(),
417
               ocl_convertScaleAbs(_src, _dst, alpha, beta))
418

419
    Mat src = _src.getMat();
420
    int cn = src.channels();
421
    double scale[] = {alpha, beta};
422
    _dst.create( src.dims, src.size, CV_8UC(cn) );
423
    Mat dst = _dst.getMat();
424
    BinaryFunc func = getCvtScaleAbsFunc(src.depth());
425
    CV_Assert( func != 0 );
426

427
    if( src.dims <= 2 )
428
    {
429
        Size sz = getContinuousSize(src, dst, cn);
430
        func( src.ptr(), src.step, 0, 0, dst.ptr(), dst.step, sz, scale );
431
    }
432
    else
433
    {
434
        const Mat* arrays[] = {&src, &dst, 0};
435
        uchar* ptrs[2] = {};
436
        NAryMatIterator it(arrays, ptrs);
437
        Size sz((int)it.size*cn, 1);
438

439
        for( size_t i = 0; i < it.nplanes; i++, ++it )
440
            func( ptrs[0], 0, 0, 0, ptrs[1], 0, sz, scale );
441
    }
442
}
443

444
//==================================================================================================
445

446
namespace cv {
447

448
#ifdef HAVE_OPENCL
449

450
static bool ocl_normalize( InputArray _src, InputOutputArray _dst, InputArray _mask, int dtype,
451
                           double scale, double delta )
452
{
453
    UMat src = _src.getUMat();
454

455
    if( _mask.empty() )
456
        src.convertTo( _dst, dtype, scale, delta );
457
    else if (src.channels() <= 4)
458
    {
459
        const ocl::Device & dev = ocl::Device::getDefault();
460

461
        int stype = _src.type(), sdepth = CV_MAT_DEPTH(stype), cn = CV_MAT_CN(stype),
462
                ddepth = CV_MAT_DEPTH(dtype), wdepth = std::max(CV_32F, std::max(sdepth, ddepth)),
463
                rowsPerWI = dev.isIntel() ? 4 : 1;
464

465
        float fscale = static_cast<float>(scale), fdelta = static_cast<float>(delta);
466
        bool haveScale = std::fabs(scale - 1) > DBL_EPSILON,
467
                haveZeroScale = !(std::fabs(scale) > DBL_EPSILON),
468
                haveDelta = std::fabs(delta) > DBL_EPSILON,
469
                doubleSupport = dev.doubleFPConfig() > 0;
470

471
        if (!haveScale && !haveDelta && stype == dtype)
472
        {
473
            _src.copyTo(_dst, _mask);
474
            return true;
475
        }
476
        if (haveZeroScale)
477
        {
478
            _dst.setTo(Scalar(delta), _mask);
479
            return true;
480
        }
481

482
        if ((sdepth == CV_64F || ddepth == CV_64F) && !doubleSupport)
483
            return false;
484

485
        char cvt[2][40];
486
        String opts = format("-D srcT=%s -D dstT=%s -D convertToWT=%s -D cn=%d -D rowsPerWI=%d"
487
                             " -D convertToDT=%s -D workT=%s%s%s%s -D srcT1=%s -D dstT1=%s",
488
                             ocl::typeToStr(stype), ocl::typeToStr(dtype),
489
                             ocl::convertTypeStr(sdepth, wdepth, cn, cvt[0]), cn,
490
                             rowsPerWI, ocl::convertTypeStr(wdepth, ddepth, cn, cvt[1]),
491
                             ocl::typeToStr(CV_MAKE_TYPE(wdepth, cn)),
492
                             doubleSupport ? " -D DOUBLE_SUPPORT" : "",
493
                             haveScale ? " -D HAVE_SCALE" : "",
494
                             haveDelta ? " -D HAVE_DELTA" : "",
495
                             ocl::typeToStr(sdepth), ocl::typeToStr(ddepth));
496

497
        ocl::Kernel k("normalizek", ocl::core::normalize_oclsrc, opts);
498
        if (k.empty())
499
            return false;
500

501
        UMat mask = _mask.getUMat(), dst = _dst.getUMat();
502

503
        ocl::KernelArg srcarg = ocl::KernelArg::ReadOnlyNoSize(src),
504
                maskarg = ocl::KernelArg::ReadOnlyNoSize(mask),
505
                dstarg = ocl::KernelArg::ReadWrite(dst);
506

507
        if (haveScale)
508
        {
509
            if (haveDelta)
510
                k.args(srcarg, maskarg, dstarg, fscale, fdelta);
511
            else
512
                k.args(srcarg, maskarg, dstarg, fscale);
513
        }
514
        else
515
        {
516
            if (haveDelta)
517
                k.args(srcarg, maskarg, dstarg, fdelta);
518
            else
519
                k.args(srcarg, maskarg, dstarg);
520
        }
521

522
        size_t globalsize[2] = { (size_t)src.cols, ((size_t)src.rows + rowsPerWI - 1) / rowsPerWI };
523
        return k.run(2, globalsize, NULL, false);
524
    }
525
    else
526
    {
527
        UMat temp;
528
        src.convertTo( temp, dtype, scale, delta );
529
        temp.copyTo( _dst, _mask );
530
    }
531

532
    return true;
533
}
534

535
#endif
536

537
} // cv::
538

539
void cv::normalize( InputArray _src, InputOutputArray _dst, double a, double b,
540
                    int norm_type, int rtype, InputArray _mask )
541
{
542
    CV_INSTRUMENT_REGION();
543

544
    double scale = 1, shift = 0;
545
    int type = _src.type(), depth = CV_MAT_DEPTH(type);
546

547
    if( rtype < 0 )
548
        rtype = _dst.fixedType() ? _dst.depth() : depth;
549

550
    if( norm_type == CV_MINMAX )
551
    {
552
        double smin = 0, smax = 0;
553
        double dmin = MIN( a, b ), dmax = MAX( a, b );
554
        minMaxIdx( _src, &smin, &smax, 0, 0, _mask );
555
        scale = (dmax - dmin)*(smax - smin > DBL_EPSILON ? 1./(smax - smin) : 0);
556
        if( rtype == CV_32F )
557
        {
558
            scale = (float)scale;
559
            shift = (float)dmin - (float)(smin*scale);
560
        }
561
        else
562
            shift = dmin - smin*scale;
563
    }
564
    else if( norm_type == CV_L2 || norm_type == CV_L1 || norm_type == CV_C )
565
    {
566
        scale = norm( _src, norm_type, _mask );
567
        scale = scale > DBL_EPSILON ? a/scale : 0.;
568
        shift = 0;
569
    }
570
    else
571
        CV_Error( CV_StsBadArg, "Unknown/unsupported norm type" );
572

573
    CV_OCL_RUN(_dst.isUMat(),
574
               ocl_normalize(_src, _dst, _mask, rtype, scale, shift))
575

576
    Mat src = _src.getMat();
577
    if( _mask.empty() )
578
        src.convertTo( _dst, rtype, scale, shift );
579
    else
580
    {
581
        Mat temp;
582
        src.convertTo( temp, rtype, scale, shift );
583
        temp.copyTo( _dst, _mask );
584
    }
585
}
586

587
Product

Resources

Company
