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 template<typename T, typename VecT> static void
13 vecsplit_( const T* src, T** dst, int len, int cn )
19 const int VECSZ = VecT::nlanes;
22 for( i = 0; i < len; i += VECSZ )
24 i = std::min( len - VECSZ, i );
26 v_load_deinterleave(src + i*cn, a, b);
34 for( i = 0; i < len; i += VECSZ )
36 i = std::min( len - VECSZ, i );
38 v_load_deinterleave(src + i*cn, a, b, c);
49 for( i = 0; i < len; i += VECSZ )
51 i = std::min( len - VECSZ, i );
53 v_load_deinterleave(src + i*cn, a, b, c, d);
64 template<typename T> static void
65 split_( const T* src, T** dst, int len, int cn )
67 int k = cn % 4 ? cn % 4 : 4;
75 memcpy(dst0, src, len * sizeof(T));
79 for( i = 0, j = 0 ; i < len; i++, j += cn )
85 T *dst0 = dst[0], *dst1 = dst[1];
88 for( ; i < len; i++, j += cn )
96 T *dst0 = dst[0], *dst1 = dst[1], *dst2 = dst[2];
99 for( ; i < len; i++, j += cn )
108 T *dst0 = dst[0], *dst1 = dst[1], *dst2 = dst[2], *dst3 = dst[3];
111 for( ; i < len; i++, j += cn )
113 dst0[i] = src[j]; dst1[i] = src[j+1];
114 dst2[i] = src[j+2]; dst3[i] = src[j+3];
118 for( ; k < cn; k += 4 )
120 T *dst0 = dst[k], *dst1 = dst[k+1], *dst2 = dst[k+2], *dst3 = dst[k+3];
121 for( i = 0, j = k; i < len; i++, j += cn )
123 dst0[i] = src[j]; dst1[i] = src[j+1];
124 dst2[i] = src[j+2]; dst3[i] = src[j+3];
129 void split8u(const uchar* src, uchar** dst, int len, int cn )
131 CALL_HAL(split8u, cv_hal_split8u, src,dst, len, cn)
134 if( len >= v_uint8::nlanes && 2 <= cn && cn <= 4 )
135 vecsplit_<uchar, v_uint8>(src, dst, len, cn);
138 split_(src, dst, len, cn);
141 void split16u(const ushort* src, ushort** dst, int len, int cn )
143 CALL_HAL(split16u, cv_hal_split16u, src,dst, len, cn)
145 if( len >= v_uint16::nlanes && 2 <= cn && cn <= 4 )
146 vecsplit_<ushort, v_uint16>(src, dst, len, cn);
149 split_(src, dst, len, cn);
152 void split32s(const int* src, int** dst, int len, int cn )
154 CALL_HAL(split32s, cv_hal_split32s, src,dst, len, cn)
156 if( len >= v_uint32::nlanes && 2 <= cn && cn <= 4 )
157 vecsplit_<int, v_int32>(src, dst, len, cn);
160 split_(src, dst, len, cn);
163 void split64s(const int64* src, int64** dst, int len, int cn )
165 CALL_HAL(split64s, cv_hal_split64s, src,dst, len, cn)
167 if( len >= v_int64::nlanes && 2 <= cn && cn <= 4 )
168 vecsplit_<int64, v_int64>(src, dst, len, cn);
171 split_(src, dst, len, cn);
176 /****************************************************************************************\
178 \****************************************************************************************/
180 typedef void (*SplitFunc)(const uchar* src, uchar** dst, int len, int cn);
182 static SplitFunc getSplitFunc(int depth)
184 static SplitFunc splitTab[] =
186 (SplitFunc)GET_OPTIMIZED(cv::hal::split8u), (SplitFunc)GET_OPTIMIZED(cv::hal::split8u), (SplitFunc)GET_OPTIMIZED(cv::hal::split16u), (SplitFunc)GET_OPTIMIZED(cv::hal::split16u),
187 (SplitFunc)GET_OPTIMIZED(cv::hal::split32s), (SplitFunc)GET_OPTIMIZED(cv::hal::split32s), (SplitFunc)GET_OPTIMIZED(cv::hal::split64s), 0
190 return splitTab[depth];
196 static bool ipp_split(const Mat& src, Mat* mv, int channels)
199 CV_INSTRUMENT_REGION_IPP()
201 if(channels != 3 && channels != 4)
206 IppiSize size = ippiSize(src.size());
207 void *dstPtrs[4] = {NULL};
208 size_t dstStep = mv[0].step;
209 for(int i = 0; i < channels; i++)
211 dstPtrs[i] = mv[i].ptr();
212 if(dstStep != mv[i].step)
216 return CV_INSTRUMENT_FUN_IPP(llwiCopySplit, src.ptr(), (int)src.step, dstPtrs, (int)dstStep, size, (int)src.elemSize1(), channels, 0) >= 0;
220 const Mat *arrays[5] = {NULL};
221 uchar *ptrs[5] = {NULL};
224 for(int i = 1; i < channels; i++)
226 arrays[i] = &mv[i-1];
229 NAryMatIterator it(arrays, ptrs);
230 IppiSize size = { (int)it.size, 1 };
232 for( size_t i = 0; i < it.nplanes; i++, ++it )
234 if(CV_INSTRUMENT_FUN_IPP(llwiCopySplit, ptrs[0], 0, (void**)&ptrs[1], 0, size, (int)src.elemSize1(), channels, 0) < 0)
240 CV_UNUSED(src); CV_UNUSED(mv); CV_UNUSED(channels);
247 void cv::split(const Mat& src, Mat* mv)
249 CV_INSTRUMENT_REGION()
251 int k, depth = src.depth(), cn = src.channels();
258 for( k = 0; k < cn; k++ )
260 mv[k].create(src.dims, src.size, depth);
263 CV_IPP_RUN_FAST(ipp_split(src, mv, cn));
265 SplitFunc func = getSplitFunc(depth);
266 CV_Assert( func != 0 );
268 size_t esz = src.elemSize(), esz1 = src.elemSize1();
269 size_t blocksize0 = (BLOCK_SIZE + esz-1)/esz;
270 AutoBuffer<uchar> _buf((cn+1)*(sizeof(Mat*) + sizeof(uchar*)) + 16);
271 const Mat** arrays = (const Mat**)_buf.data();
272 uchar** ptrs = (uchar**)alignPtr(arrays + cn + 1, 16);
275 for( k = 0; k < cn; k++ )
277 arrays[k+1] = &mv[k];
280 NAryMatIterator it(arrays, ptrs, cn+1);
281 size_t total = it.size;
282 size_t blocksize = std::min((size_t)CV_SPLIT_MERGE_MAX_BLOCK_SIZE(cn), cn <= 4 ? total : std::min(total, blocksize0));
284 for( size_t i = 0; i < it.nplanes; i++, ++it )
286 for( size_t j = 0; j < total; j += blocksize )
288 size_t bsz = std::min(total - j, blocksize);
289 func( ptrs[0], &ptrs[1], (int)bsz, cn );
291 if( j + blocksize < total )
294 for( k = 0; k < cn; k++ )
295 ptrs[k+1] += bsz*esz1;
305 static bool ocl_split( InputArray _m, OutputArrayOfArrays _mv )
307 int type = _m.type(), depth = CV_MAT_DEPTH(type), cn = CV_MAT_CN(type),
308 rowsPerWI = ocl::Device::getDefault().isIntel() ? 4 : 1;
310 String dstargs, processelem, indexdecl;
311 for (int i = 0; i < cn; ++i)
313 dstargs += format("DECLARE_DST_PARAM(%d)", i);
314 indexdecl += format("DECLARE_INDEX(%d)", i);
315 processelem += format("PROCESS_ELEM(%d)", i);
318 ocl::Kernel k("split", ocl::core::split_merge_oclsrc,
319 format("-D T=%s -D OP_SPLIT -D cn=%d -D DECLARE_DST_PARAMS=%s"
320 " -D PROCESS_ELEMS_N=%s -D DECLARE_INDEX_N=%s",
321 ocl::memopTypeToStr(depth), cn, dstargs.c_str(),
322 processelem.c_str(), indexdecl.c_str()));
326 Size size = _m.size();
327 _mv.create(cn, 1, depth);
328 for (int i = 0; i < cn; ++i)
329 _mv.create(size, depth, i);
331 std::vector<UMat> dst;
332 _mv.getUMatVector(dst);
334 int argidx = k.set(0, ocl::KernelArg::ReadOnly(_m.getUMat()));
335 for (int i = 0; i < cn; ++i)
336 argidx = k.set(argidx, ocl::KernelArg::WriteOnlyNoSize(dst[i]));
337 k.set(argidx, rowsPerWI);
339 size_t globalsize[2] = { (size_t)size.width, ((size_t)size.height + rowsPerWI - 1) / rowsPerWI };
340 return k.run(2, globalsize, NULL, false);
347 void cv::split(InputArray _m, OutputArrayOfArrays _mv)
349 CV_INSTRUMENT_REGION()
351 CV_OCL_RUN(_m.dims() <= 2 && _mv.isUMatVector(),
361 CV_Assert( !_mv.fixedType() || _mv.empty() || _mv.type() == m.depth() );
363 int depth = m.depth(), cn = m.channels();
364 _mv.create(cn, 1, depth);
365 for (int i = 0; i < cn; ++i)
366 _mv.create(m.dims, m.size.p, depth, i);
368 std::vector<Mat> dst;
369 _mv.getMatVector(dst);