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 {
13 The trick with STORE_UNALIGNED/STORE_ALIGNED_NOCACHE is the following:
14 on IA there are instructions movntps and such to which
15 v_store_interleave(...., STORE_ALIGNED_NOCACHE) is mapped.
16 Those instructions write directly into memory w/o touching cache
17 that results in dramatic speed improvements, especially on
18 large arrays (FullHD, 4K etc.).
20 Those intrinsics require the destination address to be aligned
21 by 16/32 bits (with SSE2 and AVX2, respectively).
22 So we potentially split the processing into 3 stages:
23 1) the optional prefix part [0:i0), where we use simple unaligned stores.
24 2) the optional main part [i0:len - VECSZ], where we use "nocache" mode.
25 But in some cases we have to use unaligned stores in this part.
26 3) the optional suffix part (the tail) (len - VECSZ:len) where we switch back to "unaligned" mode
27 to process the remaining len - VECSZ elements.
28 In principle there can be very poorly aligned data where there is no main part.
29 For that we set i0=0 and use unaligned stores for the whole array.
31 template<typename T, typename VecT> static void
32 vecmerge_( const T** src, T* dst, int len, int cn )
34 const int VECSZ = VecT::nlanes;
36 const T* src0 = src[0];
37 const T* src1 = src[1];
39 int r = (int)((size_t)(void*)dst % (VECSZ*sizeof(T)));
40 hal::StoreMode mode = hal::STORE_ALIGNED_NOCACHE;
43 mode = hal::STORE_UNALIGNED;
44 if( r % cn == 0 && len > VECSZ )
45 i0 = VECSZ - (r / cn);
50 for( i = 0; i < len; i += VECSZ )
55 mode = hal::STORE_UNALIGNED;
57 VecT a = vx_load(src0 + i), b = vx_load(src1 + i);
58 v_store_interleave(dst + i*cn, a, b, mode);
62 mode = hal::STORE_ALIGNED_NOCACHE;
68 const T* src2 = src[2];
69 for( i = 0; i < len; i += VECSZ )
74 mode = hal::STORE_UNALIGNED;
76 VecT a = vx_load(src0 + i), b = vx_load(src1 + i), c = vx_load(src2 + i);
77 v_store_interleave(dst + i*cn, a, b, c, mode);
81 mode = hal::STORE_ALIGNED_NOCACHE;
88 const T* src2 = src[2];
89 const T* src3 = src[3];
90 for( i = 0; i < len; i += VECSZ )
95 mode = hal::STORE_UNALIGNED;
97 VecT a = vx_load(src0 + i), b = vx_load(src1 + i);
98 VecT c = vx_load(src2 + i), d = vx_load(src3 + i);
99 v_store_interleave(dst + i*cn, a, b, c, d, mode);
103 mode = hal::STORE_ALIGNED_NOCACHE;
111 template<typename T> static void
112 merge_( const T** src, T* dst, int len, int cn )
114 int k = cn % 4 ? cn % 4 : 4;
118 const T* src0 = src[0];
119 for( i = j = 0; i < len; i++, j += cn )
124 const T *src0 = src[0], *src1 = src[1];
126 for( ; i < len; i++, j += cn )
134 const T *src0 = src[0], *src1 = src[1], *src2 = src[2];
136 for( ; i < len; i++, j += cn )
145 const T *src0 = src[0], *src1 = src[1], *src2 = src[2], *src3 = src[3];
147 for( ; i < len; i++, j += cn )
149 dst[j] = src0[i]; dst[j+1] = src1[i];
150 dst[j+2] = src2[i]; dst[j+3] = src3[i];
154 for( ; k < cn; k += 4 )
156 const T *src0 = src[k], *src1 = src[k+1], *src2 = src[k+2], *src3 = src[k+3];
157 for( i = 0, j = k; i < len; i++, j += cn )
159 dst[j] = src0[i]; dst[j+1] = src1[i];
160 dst[j+2] = src2[i]; dst[j+3] = src3[i];
165 void merge8u(const uchar** src, uchar* dst, int len, int cn )
167 CALL_HAL(merge8u, cv_hal_merge8u, src, dst, len, cn)
169 if( len >= v_uint8::nlanes && 2 <= cn && cn <= 4 )
170 vecmerge_<uchar, v_uint8>(src, dst, len, cn);
173 merge_(src, dst, len, cn);
176 void merge16u(const ushort** src, ushort* dst, int len, int cn )
178 CALL_HAL(merge16u, cv_hal_merge16u, src, dst, len, cn)
180 if( len >= v_uint16::nlanes && 2 <= cn && cn <= 4 )
181 vecmerge_<ushort, v_uint16>(src, dst, len, cn);
184 merge_(src, dst, len, cn);
187 void merge32s(const int** src, int* dst, int len, int cn )
189 CALL_HAL(merge32s, cv_hal_merge32s, src, dst, len, cn)
191 if( len >= v_int32::nlanes && 2 <= cn && cn <= 4 )
192 vecmerge_<int, v_int32>(src, dst, len, cn);
195 merge_(src, dst, len, cn);
198 void merge64s(const int64** src, int64* dst, int len, int cn )
200 CALL_HAL(merge64s, cv_hal_merge64s, src, dst, len, cn)
202 if( len >= v_int64::nlanes && 2 <= cn && cn <= 4 )
203 vecmerge_<int64, v_int64>(src, dst, len, cn);
206 merge_(src, dst, len, cn);
212 typedef void (*MergeFunc)(const uchar** src, uchar* dst, int len, int cn);
214 static MergeFunc getMergeFunc(int depth)
216 static MergeFunc mergeTab[] =
218 (MergeFunc)GET_OPTIMIZED(cv::hal::merge8u), (MergeFunc)GET_OPTIMIZED(cv::hal::merge8u), (MergeFunc)GET_OPTIMIZED(cv::hal::merge16u), (MergeFunc)GET_OPTIMIZED(cv::hal::merge16u),
219 (MergeFunc)GET_OPTIMIZED(cv::hal::merge32s), (MergeFunc)GET_OPTIMIZED(cv::hal::merge32s), (MergeFunc)GET_OPTIMIZED(cv::hal::merge64s), 0
222 return mergeTab[depth];
228 static bool ipp_merge(const Mat* mv, Mat& dst, int channels)
231 CV_INSTRUMENT_REGION_IPP()
233 if(channels != 3 && channels != 4)
238 IppiSize size = ippiSize(mv[0].size());
239 const void *srcPtrs[4] = {NULL};
240 size_t srcStep = mv[0].step;
241 for(int i = 0; i < channels; i++)
243 srcPtrs[i] = mv[i].ptr();
244 if(srcStep != mv[i].step)
248 return CV_INSTRUMENT_FUN_IPP(llwiCopyMerge, srcPtrs, (int)srcStep, dst.ptr(), (int)dst.step, size, (int)mv[0].elemSize1(), channels, 0) >= 0;
252 const Mat *arrays[5] = {NULL};
253 uchar *ptrs[5] = {NULL};
256 for(int i = 1; i < channels; i++)
258 arrays[i] = &mv[i-1];
261 NAryMatIterator it(arrays, ptrs);
262 IppiSize size = { (int)it.size, 1 };
264 for( size_t i = 0; i < it.nplanes; i++, ++it )
266 if(CV_INSTRUMENT_FUN_IPP(llwiCopyMerge, (const void**)&ptrs[1], 0, ptrs[0], 0, size, (int)mv[0].elemSize1(), channels, 0) < 0)
272 CV_UNUSED(dst); CV_UNUSED(mv); CV_UNUSED(channels);
279 void cv::merge(const Mat* mv, size_t n, OutputArray _dst)
281 CV_INSTRUMENT_REGION()
283 CV_Assert( mv && n > 0 );
285 int depth = mv[0].depth();
290 for( i = 0; i < n; i++ )
292 CV_Assert(mv[i].size == mv[0].size && mv[i].depth() == depth);
293 allch1 = allch1 && mv[i].channels() == 1;
294 cn += mv[i].channels();
297 CV_Assert( 0 < cn && cn <= CV_CN_MAX );
298 _dst.create(mv[0].dims, mv[0].size, CV_MAKETYPE(depth, cn));
299 Mat dst = _dst.getMat();
307 CV_IPP_RUN_FAST(ipp_merge(mv, dst, (int)n));
311 AutoBuffer<int> pairs(cn*2);
314 for( i = 0, j = 0; i < n; i++, j += ni )
316 ni = mv[i].channels();
317 for( k = 0; k < ni; k++ )
319 pairs[(j+k)*2] = j + k;
320 pairs[(j+k)*2+1] = j + k;
323 mixChannels( mv, n, &dst, 1, &pairs[0], cn );
327 MergeFunc func = getMergeFunc(depth);
328 CV_Assert( func != 0 );
330 size_t esz = dst.elemSize(), esz1 = dst.elemSize1();
331 size_t blocksize0 = (int)((BLOCK_SIZE + esz-1)/esz);
332 AutoBuffer<uchar> _buf((cn+1)*(sizeof(Mat*) + sizeof(uchar*)) + 16);
333 const Mat** arrays = (const Mat**)_buf.data();
334 uchar** ptrs = (uchar**)alignPtr(arrays + cn + 1, 16);
337 for( k = 0; k < cn; k++ )
338 arrays[k+1] = &mv[k];
340 NAryMatIterator it(arrays, ptrs, cn+1);
341 size_t total = (int)it.size;
342 size_t blocksize = std::min((size_t)CV_SPLIT_MERGE_MAX_BLOCK_SIZE(cn), cn <= 4 ? total : std::min(total, blocksize0));
344 for( i = 0; i < it.nplanes; i++, ++it )
346 for( size_t j = 0; j < total; j += blocksize )
348 size_t bsz = std::min(total - j, blocksize);
349 func( (const uchar**)&ptrs[1], ptrs[0], (int)bsz, cn );
351 if( j + blocksize < total )
354 for( int t = 0; t < cn; t++ )
355 ptrs[t+1] += bsz*esz1;
365 static bool ocl_merge( InputArrayOfArrays _mv, OutputArray _dst )
367 std::vector<UMat> src, ksrc;
368 _mv.getUMatVector(src);
369 CV_Assert(!src.empty());
371 int type = src[0].type(), depth = CV_MAT_DEPTH(type),
372 rowsPerWI = ocl::Device::getDefault().isIntel() ? 4 : 1;
373 Size size = src[0].size();
375 for (size_t i = 0, srcsize = src.size(); i < srcsize; ++i)
377 int itype = src[i].type(), icn = CV_MAT_CN(itype), idepth = CV_MAT_DEPTH(itype),
378 esz1 = CV_ELEM_SIZE1(idepth);
382 CV_Assert(size == src[i].size() && depth == idepth);
384 for (int cn = 0; cn < icn; ++cn)
387 tsrc.offset += cn * esz1;
388 ksrc.push_back(tsrc);
391 int dcn = (int)ksrc.size();
393 String srcargs, processelem, cndecl, indexdecl;
394 for (int i = 0; i < dcn; ++i)
396 srcargs += format("DECLARE_SRC_PARAM(%d)", i);
397 processelem += format("PROCESS_ELEM(%d)", i);
398 indexdecl += format("DECLARE_INDEX(%d)", i);
399 cndecl += format(" -D scn%d=%d", i, ksrc[i].channels());
402 ocl::Kernel k("merge", ocl::core::split_merge_oclsrc,
403 format("-D OP_MERGE -D cn=%d -D T=%s -D DECLARE_SRC_PARAMS_N=%s"
404 " -D DECLARE_INDEX_N=%s -D PROCESS_ELEMS_N=%s%s",
405 dcn, ocl::memopTypeToStr(depth), srcargs.c_str(),
406 indexdecl.c_str(), processelem.c_str(), cndecl.c_str()));
410 _dst.create(size, CV_MAKE_TYPE(depth, dcn));
411 UMat dst = _dst.getUMat();
414 for (int i = 0; i < dcn; ++i)
415 argidx = k.set(argidx, ocl::KernelArg::ReadOnlyNoSize(ksrc[i]));
416 argidx = k.set(argidx, ocl::KernelArg::WriteOnly(dst));
417 k.set(argidx, rowsPerWI);
419 size_t globalsize[2] = { (size_t)dst.cols, ((size_t)dst.rows + rowsPerWI - 1) / rowsPerWI };
420 return k.run(2, globalsize, NULL, false);
427 void cv::merge(InputArrayOfArrays _mv, OutputArray _dst)
429 CV_INSTRUMENT_REGION()
431 CV_OCL_RUN(_mv.isUMatVector() && _dst.isUMat(),
432 ocl_merge(_mv, _dst))
435 _mv.getMatVector(mv);
436 merge(!mv.empty() ? &mv[0] : 0, mv.size(), _dst);