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
7 #include "opencl_kernels_core.hpp"
9 namespace cv { namespace hal {
12 // see the comments for vecmerge_ in merge.cpp
13 template<typename T, typename VecT> static void
14 vecsplit_( const T* src, T** dst, int len, int cn )
16 const int VECSZ = VecT::nlanes;
21 int r0 = (int)((size_t)(void*)dst0 % (VECSZ*sizeof(T)));
22 int r1 = (int)((size_t)(void*)dst1 % (VECSZ*sizeof(T)));
23 int r2 = cn > 2 ? (int)((size_t)(void*)dst[2] % (VECSZ*sizeof(T))) : r0;
24 int r3 = cn > 3 ? (int)((size_t)(void*)dst[3] % (VECSZ*sizeof(T))) : r0;
26 hal::StoreMode mode = hal::STORE_ALIGNED_NOCACHE;
27 if( (r0|r1|r2|r3) != 0 )
29 mode = hal::STORE_UNALIGNED;
30 if( r0 == r1 && r0 == r2 && r0 == r3 && r0 % cn == 0 && len > VECSZ )
31 i0 = VECSZ - (r0 / cn);
36 for( i = 0; i < len; i += VECSZ )
41 mode = hal::STORE_UNALIGNED;
44 v_load_deinterleave(src + i*cn, a, b);
45 v_store(dst0 + i, a, mode);
46 v_store(dst1 + i, b, mode);
50 mode = hal::STORE_ALIGNED_NOCACHE;
57 for( i = 0; i < len; i += VECSZ )
62 mode = hal::STORE_UNALIGNED;
65 v_load_deinterleave(src + i*cn, a, b, c);
66 v_store(dst0 + i, a, mode);
67 v_store(dst1 + i, b, mode);
68 v_store(dst2 + i, c, mode);
72 mode = hal::STORE_ALIGNED_NOCACHE;
81 for( i = 0; i < len; i += VECSZ )
86 mode = hal::STORE_UNALIGNED;
89 v_load_deinterleave(src + i*cn, a, b, c, d);
90 v_store(dst0 + i, a, mode);
91 v_store(dst1 + i, b, mode);
92 v_store(dst2 + i, c, mode);
93 v_store(dst3 + i, d, mode);
97 mode = hal::STORE_ALIGNED_NOCACHE;
105 template<typename T> static void
106 split_( const T* src, T** dst, int len, int cn )
108 int k = cn % 4 ? cn % 4 : 4;
116 memcpy(dst0, src, len * sizeof(T));
120 for( i = 0, j = 0 ; i < len; i++, j += cn )
126 T *dst0 = dst[0], *dst1 = dst[1];
129 for( ; i < len; i++, j += cn )
137 T *dst0 = dst[0], *dst1 = dst[1], *dst2 = dst[2];
140 for( ; i < len; i++, j += cn )
149 T *dst0 = dst[0], *dst1 = dst[1], *dst2 = dst[2], *dst3 = dst[3];
152 for( ; i < len; i++, j += cn )
154 dst0[i] = src[j]; dst1[i] = src[j+1];
155 dst2[i] = src[j+2]; dst3[i] = src[j+3];
159 for( ; k < cn; k += 4 )
161 T *dst0 = dst[k], *dst1 = dst[k+1], *dst2 = dst[k+2], *dst3 = dst[k+3];
162 for( i = 0, j = k; i < len; i++, j += cn )
164 dst0[i] = src[j]; dst1[i] = src[j+1];
165 dst2[i] = src[j+2]; dst3[i] = src[j+3];
170 void split8u(const uchar* src, uchar** dst, int len, int cn )
172 CALL_HAL(split8u, cv_hal_split8u, src,dst, len, cn)
175 if( len >= v_uint8::nlanes && 2 <= cn && cn <= 4 )
176 vecsplit_<uchar, v_uint8>(src, dst, len, cn);
179 split_(src, dst, len, cn);
182 void split16u(const ushort* src, ushort** dst, int len, int cn )
184 CALL_HAL(split16u, cv_hal_split16u, src,dst, len, cn)
186 if( len >= v_uint16::nlanes && 2 <= cn && cn <= 4 )
187 vecsplit_<ushort, v_uint16>(src, dst, len, cn);
190 split_(src, dst, len, cn);
193 void split32s(const int* src, int** dst, int len, int cn )
195 CALL_HAL(split32s, cv_hal_split32s, src,dst, len, cn)
197 if( len >= v_uint32::nlanes && 2 <= cn && cn <= 4 )
198 vecsplit_<int, v_int32>(src, dst, len, cn);
201 split_(src, dst, len, cn);
204 void split64s(const int64* src, int64** dst, int len, int cn )
206 CALL_HAL(split64s, cv_hal_split64s, src,dst, len, cn)
208 if( len >= v_int64::nlanes && 2 <= cn && cn <= 4 )
209 vecsplit_<int64, v_int64>(src, dst, len, cn);
212 split_(src, dst, len, cn);
217 /****************************************************************************************\
219 \****************************************************************************************/
221 typedef void (*SplitFunc)(const uchar* src, uchar** dst, int len, int cn);
223 static SplitFunc getSplitFunc(int depth)
225 static SplitFunc splitTab[] =
227 (SplitFunc)GET_OPTIMIZED(cv::hal::split8u), (SplitFunc)GET_OPTIMIZED(cv::hal::split8u), (SplitFunc)GET_OPTIMIZED(cv::hal::split16u), (SplitFunc)GET_OPTIMIZED(cv::hal::split16u),
228 (SplitFunc)GET_OPTIMIZED(cv::hal::split32s), (SplitFunc)GET_OPTIMIZED(cv::hal::split32s), (SplitFunc)GET_OPTIMIZED(cv::hal::split64s), 0
231 return splitTab[depth];
237 static bool ipp_split(const Mat& src, Mat* mv, int channels)
240 CV_INSTRUMENT_REGION_IPP()
242 if(channels != 3 && channels != 4)
247 IppiSize size = ippiSize(src.size());
248 void *dstPtrs[4] = {NULL};
249 size_t dstStep = mv[0].step;
250 for(int i = 0; i < channels; i++)
252 dstPtrs[i] = mv[i].ptr();
253 if(dstStep != mv[i].step)
257 return CV_INSTRUMENT_FUN_IPP(llwiCopySplit, src.ptr(), (int)src.step, dstPtrs, (int)dstStep, size, (int)src.elemSize1(), channels, 0) >= 0;
261 const Mat *arrays[5] = {NULL};
262 uchar *ptrs[5] = {NULL};
265 for(int i = 1; i < channels; i++)
267 arrays[i] = &mv[i-1];
270 NAryMatIterator it(arrays, ptrs);
271 IppiSize size = { (int)it.size, 1 };
273 for( size_t i = 0; i < it.nplanes; i++, ++it )
275 if(CV_INSTRUMENT_FUN_IPP(llwiCopySplit, ptrs[0], 0, (void**)&ptrs[1], 0, size, (int)src.elemSize1(), channels, 0) < 0)
281 CV_UNUSED(src); CV_UNUSED(mv); CV_UNUSED(channels);
288 void cv::split(const Mat& src, Mat* mv)
290 CV_INSTRUMENT_REGION()
292 int k, depth = src.depth(), cn = src.channels();
299 for( k = 0; k < cn; k++ )
301 mv[k].create(src.dims, src.size, depth);
304 CV_IPP_RUN_FAST(ipp_split(src, mv, cn));
306 SplitFunc func = getSplitFunc(depth);
307 CV_Assert( func != 0 );
309 size_t esz = src.elemSize(), esz1 = src.elemSize1();
310 size_t blocksize0 = (BLOCK_SIZE + esz-1)/esz;
311 AutoBuffer<uchar> _buf((cn+1)*(sizeof(Mat*) + sizeof(uchar*)) + 16);
312 const Mat** arrays = (const Mat**)_buf.data();
313 uchar** ptrs = (uchar**)alignPtr(arrays + cn + 1, 16);
316 for( k = 0; k < cn; k++ )
318 arrays[k+1] = &mv[k];
321 NAryMatIterator it(arrays, ptrs, cn+1);
322 size_t total = it.size;
323 size_t blocksize = std::min((size_t)CV_SPLIT_MERGE_MAX_BLOCK_SIZE(cn), cn <= 4 ? total : std::min(total, blocksize0));
325 for( size_t i = 0; i < it.nplanes; i++, ++it )
327 for( size_t j = 0; j < total; j += blocksize )
329 size_t bsz = std::min(total - j, blocksize);
330 func( ptrs[0], &ptrs[1], (int)bsz, cn );
332 if( j + blocksize < total )
335 for( k = 0; k < cn; k++ )
336 ptrs[k+1] += bsz*esz1;
346 static bool ocl_split( InputArray _m, OutputArrayOfArrays _mv )
348 int type = _m.type(), depth = CV_MAT_DEPTH(type), cn = CV_MAT_CN(type),
349 rowsPerWI = ocl::Device::getDefault().isIntel() ? 4 : 1;
351 String dstargs, processelem, indexdecl;
352 for (int i = 0; i < cn; ++i)
354 dstargs += format("DECLARE_DST_PARAM(%d)", i);
355 indexdecl += format("DECLARE_INDEX(%d)", i);
356 processelem += format("PROCESS_ELEM(%d)", i);
359 ocl::Kernel k("split", ocl::core::split_merge_oclsrc,
360 format("-D T=%s -D OP_SPLIT -D cn=%d -D DECLARE_DST_PARAMS=%s"
361 " -D PROCESS_ELEMS_N=%s -D DECLARE_INDEX_N=%s",
362 ocl::memopTypeToStr(depth), cn, dstargs.c_str(),
363 processelem.c_str(), indexdecl.c_str()));
367 Size size = _m.size();
368 _mv.create(cn, 1, depth);
369 for (int i = 0; i < cn; ++i)
370 _mv.create(size, depth, i);
372 std::vector<UMat> dst;
373 _mv.getUMatVector(dst);
375 int argidx = k.set(0, ocl::KernelArg::ReadOnly(_m.getUMat()));
376 for (int i = 0; i < cn; ++i)
377 argidx = k.set(argidx, ocl::KernelArg::WriteOnlyNoSize(dst[i]));
378 k.set(argidx, rowsPerWI);
380 size_t globalsize[2] = { (size_t)size.width, ((size_t)size.height + rowsPerWI - 1) / rowsPerWI };
381 return k.run(2, globalsize, NULL, false);
388 void cv::split(InputArray _m, OutputArrayOfArrays _mv)
390 CV_INSTRUMENT_REGION()
392 CV_OCL_RUN(_m.dims() <= 2 && _mv.isUMatVector(),
402 CV_Assert( !_mv.fixedType() || _mv.empty() || _mv.type() == m.depth() );
404 int depth = m.depth(), cn = m.channels();
405 _mv.create(cn, 1, depth);
406 for (int i = 0; i < cn; ++i)
407 _mv.create(m.dims, m.size.p, depth, i);
409 std::vector<Mat> dst;
410 _mv.getMatVector(dst);