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 vecmerge_( const T** src, T* dst, int len, int cn )
16 const T* src0 = src[0];
17 const T* src1 = src[1];
19 const int VECSZ = VecT::nlanes;
22 for( i = 0; i < len; i += VECSZ )
24 i = std::min( len - VECSZ, i );
25 VecT a = vx_load(src0 + i), b = vx_load(src1 + i);
26 v_store_interleave(dst + i*cn, a, b);
31 const T* src2 = src[2];
32 for( i = 0; i < len; i += VECSZ )
34 i = std::min( len - VECSZ, i );
35 VecT a = vx_load(src0 + i), b = vx_load(src1 + i), c = vx_load(src2 + i);
36 v_store_interleave(dst + i*cn, a, b, c);
42 const T* src2 = src[2];
43 const T* src3 = src[3];
44 for( i = 0; i < len; i += VECSZ )
46 i = std::min( len - VECSZ, i );
47 VecT a = vx_load(src0 + i), b = vx_load(src1 + i);
48 VecT c = vx_load(src2 + i), d = vx_load(src3 + i);
49 v_store_interleave(dst + i*cn, a, b, c, d);
56 template<typename T> static void
57 merge_( const T** src, T* dst, int len, int cn )
59 int k = cn % 4 ? cn % 4 : 4;
63 const T* src0 = src[0];
64 for( i = j = 0; i < len; i++, j += cn )
69 const T *src0 = src[0], *src1 = src[1];
71 for( ; i < len; i++, j += cn )
79 const T *src0 = src[0], *src1 = src[1], *src2 = src[2];
81 for( ; i < len; i++, j += cn )
90 const T *src0 = src[0], *src1 = src[1], *src2 = src[2], *src3 = src[3];
92 for( ; i < len; i++, j += cn )
94 dst[j] = src0[i]; dst[j+1] = src1[i];
95 dst[j+2] = src2[i]; dst[j+3] = src3[i];
99 for( ; k < cn; k += 4 )
101 const T *src0 = src[k], *src1 = src[k+1], *src2 = src[k+2], *src3 = src[k+3];
102 for( i = 0, j = k; i < len; i++, j += cn )
104 dst[j] = src0[i]; dst[j+1] = src1[i];
105 dst[j+2] = src2[i]; dst[j+3] = src3[i];
110 void merge8u(const uchar** src, uchar* dst, int len, int cn )
112 CALL_HAL(merge8u, cv_hal_merge8u, src, dst, len, cn)
114 if( len >= v_uint8::nlanes && 2 <= cn && cn <= 4 )
115 vecmerge_<uchar, v_uint8>(src, dst, len, cn);
118 merge_(src, dst, len, cn);
121 void merge16u(const ushort** src, ushort* dst, int len, int cn )
123 CALL_HAL(merge16u, cv_hal_merge16u, src, dst, len, cn)
125 if( len >= v_uint16::nlanes && 2 <= cn && cn <= 4 )
126 vecmerge_<ushort, v_uint16>(src, dst, len, cn);
129 merge_(src, dst, len, cn);
132 void merge32s(const int** src, int* dst, int len, int cn )
134 CALL_HAL(merge32s, cv_hal_merge32s, src, dst, len, cn)
136 if( len >= v_int32::nlanes && 2 <= cn && cn <= 4 )
137 vecmerge_<int, v_int32>(src, dst, len, cn);
140 merge_(src, dst, len, cn);
143 void merge64s(const int64** src, int64* dst, int len, int cn )
145 CALL_HAL(merge64s, cv_hal_merge64s, src, dst, len, cn)
147 if( len >= v_int64::nlanes && 2 <= cn && cn <= 4 )
148 vecmerge_<int64, v_int64>(src, dst, len, cn);
151 merge_(src, dst, len, cn);
157 typedef void (*MergeFunc)(const uchar** src, uchar* dst, int len, int cn);
159 static MergeFunc getMergeFunc(int depth)
161 static MergeFunc mergeTab[] =
163 (MergeFunc)GET_OPTIMIZED(cv::hal::merge8u), (MergeFunc)GET_OPTIMIZED(cv::hal::merge8u), (MergeFunc)GET_OPTIMIZED(cv::hal::merge16u), (MergeFunc)GET_OPTIMIZED(cv::hal::merge16u),
164 (MergeFunc)GET_OPTIMIZED(cv::hal::merge32s), (MergeFunc)GET_OPTIMIZED(cv::hal::merge32s), (MergeFunc)GET_OPTIMIZED(cv::hal::merge64s), 0
167 return mergeTab[depth];
173 static bool ipp_merge(const Mat* mv, Mat& dst, int channels)
176 CV_INSTRUMENT_REGION_IPP()
178 if(channels != 3 && channels != 4)
183 IppiSize size = ippiSize(mv[0].size());
184 const void *srcPtrs[4] = {NULL};
185 size_t srcStep = mv[0].step;
186 for(int i = 0; i < channels; i++)
188 srcPtrs[i] = mv[i].ptr();
189 if(srcStep != mv[i].step)
193 return CV_INSTRUMENT_FUN_IPP(llwiCopyMerge, srcPtrs, (int)srcStep, dst.ptr(), (int)dst.step, size, (int)mv[0].elemSize1(), channels, 0) >= 0;
197 const Mat *arrays[5] = {NULL};
198 uchar *ptrs[5] = {NULL};
201 for(int i = 1; i < channels; i++)
203 arrays[i] = &mv[i-1];
206 NAryMatIterator it(arrays, ptrs);
207 IppiSize size = { (int)it.size, 1 };
209 for( size_t i = 0; i < it.nplanes; i++, ++it )
211 if(CV_INSTRUMENT_FUN_IPP(llwiCopyMerge, (const void**)&ptrs[1], 0, ptrs[0], 0, size, (int)mv[0].elemSize1(), channels, 0) < 0)
217 CV_UNUSED(dst); CV_UNUSED(mv); CV_UNUSED(channels);
224 void cv::merge(const Mat* mv, size_t n, OutputArray _dst)
226 CV_INSTRUMENT_REGION()
228 CV_Assert( mv && n > 0 );
230 int depth = mv[0].depth();
235 for( i = 0; i < n; i++ )
237 CV_Assert(mv[i].size == mv[0].size && mv[i].depth() == depth);
238 allch1 = allch1 && mv[i].channels() == 1;
239 cn += mv[i].channels();
242 CV_Assert( 0 < cn && cn <= CV_CN_MAX );
243 _dst.create(mv[0].dims, mv[0].size, CV_MAKETYPE(depth, cn));
244 Mat dst = _dst.getMat();
252 CV_IPP_RUN_FAST(ipp_merge(mv, dst, (int)n));
256 AutoBuffer<int> pairs(cn*2);
259 for( i = 0, j = 0; i < n; i++, j += ni )
261 ni = mv[i].channels();
262 for( k = 0; k < ni; k++ )
264 pairs[(j+k)*2] = j + k;
265 pairs[(j+k)*2+1] = j + k;
268 mixChannels( mv, n, &dst, 1, &pairs[0], cn );
272 MergeFunc func = getMergeFunc(depth);
273 CV_Assert( func != 0 );
275 size_t esz = dst.elemSize(), esz1 = dst.elemSize1();
276 size_t blocksize0 = (int)((BLOCK_SIZE + esz-1)/esz);
277 AutoBuffer<uchar> _buf((cn+1)*(sizeof(Mat*) + sizeof(uchar*)) + 16);
278 const Mat** arrays = (const Mat**)_buf.data();
279 uchar** ptrs = (uchar**)alignPtr(arrays + cn + 1, 16);
282 for( k = 0; k < cn; k++ )
283 arrays[k+1] = &mv[k];
285 NAryMatIterator it(arrays, ptrs, cn+1);
286 size_t total = (int)it.size;
287 size_t blocksize = std::min((size_t)CV_SPLIT_MERGE_MAX_BLOCK_SIZE(cn), cn <= 4 ? total : std::min(total, blocksize0));
289 for( i = 0; i < it.nplanes; i++, ++it )
291 for( size_t j = 0; j < total; j += blocksize )
293 size_t bsz = std::min(total - j, blocksize);
294 func( (const uchar**)&ptrs[1], ptrs[0], (int)bsz, cn );
296 if( j + blocksize < total )
299 for( int t = 0; t < cn; t++ )
300 ptrs[t+1] += bsz*esz1;
310 static bool ocl_merge( InputArrayOfArrays _mv, OutputArray _dst )
312 std::vector<UMat> src, ksrc;
313 _mv.getUMatVector(src);
314 CV_Assert(!src.empty());
316 int type = src[0].type(), depth = CV_MAT_DEPTH(type),
317 rowsPerWI = ocl::Device::getDefault().isIntel() ? 4 : 1;
318 Size size = src[0].size();
320 for (size_t i = 0, srcsize = src.size(); i < srcsize; ++i)
322 int itype = src[i].type(), icn = CV_MAT_CN(itype), idepth = CV_MAT_DEPTH(itype),
323 esz1 = CV_ELEM_SIZE1(idepth);
327 CV_Assert(size == src[i].size() && depth == idepth);
329 for (int cn = 0; cn < icn; ++cn)
332 tsrc.offset += cn * esz1;
333 ksrc.push_back(tsrc);
336 int dcn = (int)ksrc.size();
338 String srcargs, processelem, cndecl, indexdecl;
339 for (int i = 0; i < dcn; ++i)
341 srcargs += format("DECLARE_SRC_PARAM(%d)", i);
342 processelem += format("PROCESS_ELEM(%d)", i);
343 indexdecl += format("DECLARE_INDEX(%d)", i);
344 cndecl += format(" -D scn%d=%d", i, ksrc[i].channels());
347 ocl::Kernel k("merge", ocl::core::split_merge_oclsrc,
348 format("-D OP_MERGE -D cn=%d -D T=%s -D DECLARE_SRC_PARAMS_N=%s"
349 " -D DECLARE_INDEX_N=%s -D PROCESS_ELEMS_N=%s%s",
350 dcn, ocl::memopTypeToStr(depth), srcargs.c_str(),
351 indexdecl.c_str(), processelem.c_str(), cndecl.c_str()));
355 _dst.create(size, CV_MAKE_TYPE(depth, dcn));
356 UMat dst = _dst.getUMat();
359 for (int i = 0; i < dcn; ++i)
360 argidx = k.set(argidx, ocl::KernelArg::ReadOnlyNoSize(ksrc[i]));
361 argidx = k.set(argidx, ocl::KernelArg::WriteOnly(dst));
362 k.set(argidx, rowsPerWI);
364 size_t globalsize[2] = { (size_t)dst.cols, ((size_t)dst.rows + rowsPerWI - 1) / rowsPerWI };
365 return k.run(2, globalsize, NULL, false);
372 void cv::merge(InputArrayOfArrays _mv, OutputArray _dst)
374 CV_INSTRUMENT_REGION()
376 CV_OCL_RUN(_mv.isUMatVector() && _dst.isUMat(),
377 ocl_merge(_mv, _dst))
380 _mv.getMatVector(mv);
381 merge(!mv.empty() ? &mv[0] : 0, mv.size(), _dst);