1 /*M///////////////////////////////////////////////////////////////////////////////////////
3 // IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
5 // By downloading, copying, installing or using the software you agree to this license.
6 // If you do not agree to this license, do not download, install,
7 // copy or use the software.
11 // For Open Source Computer Vision Library
13 // Copyright (C) 2010-2012, Institute Of Software Chinese Academy Of Science, all rights reserved.
14 // Copyright (C) 2010-2012, Advanced Micro Devices, Inc., all rights reserved.
15 // Third party copyrights are property of their respective owners.
18 // Jia Haipeng, jiahaipeng95@gmail.com
20 // Redistribution and use in source and binary forms, with or without modification,
21 // are permitted provided that the following conditions are met:
23 // * Redistribution's of source code must retain the above copyright notice,
24 // this list of conditions and the following disclaimer.
26 // * Redistribution's in binary form must reproduce the above copyright notice,
27 // this list of conditions and the following disclaimer in the documentation
28 // and/or other oclMaterials provided with the distribution.
30 // * The name of the copyright holders may not be used to endorse or promote products
31 // derived from this software without specific prior written permission.
33 // This software is provided by the copyright holders and contributors "as is" and
34 // any express or implied warranties, including, but not limited to, the implied
35 // warranties of merchantability and fitness for a particular purpose are disclaimed.
36 // In no event shall the Intel Corporation or contributors be liable for any direct,
37 // indirect, incidental, special, exemplary, or consequential damages
38 // (including, but not limited to, procurement of substitute goods or services;
39 // loss of use, data, or profits; or business interruption) however caused
40 // and on any theory of liability, whether in contract, strict liability,
41 // or tort (including negligence or otherwise) arising in any way out of
42 // the use of this software, even if advised of the possibility of such damage.
46 #include "precomp.hpp"
50 using namespace cv::ocl;
57 ////////////////////////////////////////////////////////////////////////
58 ///////////////// oclMat merge and split ///////////////////////////////
59 ////////////////////////////////////////////////////////////////////////
65 ///////////////////////////OpenCL kernel strings///////////////////////////
66 extern const char *merge_mat;
67 extern const char *split_mat;
76 static void merge_vector_run(const oclMat *mat_src, size_t n, oclMat &mat_dst)
78 if(!mat_dst.clCxt->supportsFeature(Context::CL_DOUBLE) && mat_dst.type() == CV_64F)
80 CV_Error(CV_GpuNotSupported, "Selected device don't support double\r\n");
84 Context *clCxt = mat_dst.clCxt;
85 int channels = mat_dst.oclchannels();
86 int depth = mat_dst.depth();
88 string kernelName = "merge_vector";
90 int vector_lengths[4][7] = {{0, 0, 0, 0, 0, 0, 0},
91 {2, 2, 1, 1, 1, 1, 1},
92 {4, 4, 2, 2 , 1, 1, 1},
96 size_t vector_length = vector_lengths[channels - 1][depth];
97 int offset_cols = (mat_dst.offset / mat_dst.elemSize()) & (vector_length - 1);
98 int cols = divUp(mat_dst.cols + offset_cols, vector_length);
100 size_t localThreads[3] = { 64, 4, 1 };
101 size_t globalThreads[3] = { cols, mat_dst.rows, 1 };
103 int dst_step1 = mat_dst.cols * mat_dst.elemSize();
104 vector<pair<size_t , const void *> > args;
105 args.push_back( make_pair( sizeof(cl_mem), (void *)&mat_dst.data));
106 args.push_back( make_pair( sizeof(cl_int), (void *)&mat_dst.step));
107 args.push_back( make_pair( sizeof(cl_int), (void *)&mat_dst.offset));
108 args.push_back( make_pair( sizeof(cl_mem), (void *)&mat_src[0].data));
109 args.push_back( make_pair( sizeof(cl_int), (void *)&mat_src[0].step));
110 args.push_back( make_pair( sizeof(cl_int), (void *)&mat_src[0].offset));
111 args.push_back( make_pair( sizeof(cl_mem), (void *)&mat_src[1].data));
112 args.push_back( make_pair( sizeof(cl_int), (void *)&mat_src[1].step));
113 args.push_back( make_pair( sizeof(cl_int), (void *)&mat_src[1].offset));
117 args.push_back( make_pair( sizeof(cl_mem), (void *)&mat_src[2].data));
118 args.push_back( make_pair( sizeof(cl_int), (void *)&mat_src[2].step));
119 args.push_back( make_pair( sizeof(cl_int), (void *)&mat_src[2].offset));
123 args.push_back( make_pair( sizeof(cl_mem), (void *)&mat_src[2].data));
124 args.push_back( make_pair( sizeof(cl_int), (void *)&mat_src[2].step));
125 args.push_back( make_pair( sizeof(cl_int), (void *)&mat_src[2].offset));
129 args.push_back( make_pair( sizeof(cl_mem), (void *)&mat_src[3].data));
130 args.push_back( make_pair( sizeof(cl_int), (void *)&mat_src[3].step));
131 args.push_back( make_pair( sizeof(cl_int), (void *)&mat_src[3].offset));
135 args.push_back( make_pair( sizeof(cl_int), (void *)&mat_dst.rows));
136 args.push_back( make_pair( sizeof(cl_int), (void *)&cols));
137 args.push_back( make_pair( sizeof(cl_int), (void *)&dst_step1));
139 openCLExecuteKernel(clCxt, &merge_mat, kernelName, globalThreads, localThreads, args, channels, depth);
141 static void merge(const oclMat *mat_src, size_t n, oclMat &mat_dst)
146 int depth = mat_src[0].depth();
147 Size size = mat_src[0].size();
149 int total_channels = 0;
151 for(size_t i = 0; i < n; ++i)
153 CV_Assert(depth == mat_src[i].depth());
154 CV_Assert(size == mat_src[i].size());
156 total_channels += mat_src[i].oclchannels();
159 CV_Assert(total_channels <= 4);
161 if(total_channels == 1)
163 mat_src[0].copyTo(mat_dst);
167 mat_dst.create(size, CV_MAKETYPE(depth, total_channels));
168 merge_vector_run(mat_src, n, mat_dst);
170 static void split_vector_run(const oclMat &mat_src, oclMat *mat_dst)
173 if(!mat_src.clCxt->supportsFeature(Context::CL_DOUBLE) && mat_src.type() == CV_64F)
175 CV_Error(CV_GpuNotSupported, "Selected device don't support double\r\n");
179 Context *clCxt = mat_src.clCxt;
180 int channels = mat_src.oclchannels();
181 int depth = mat_src.depth();
183 string kernelName = "split_vector";
185 int vector_lengths[4][7] = {{0, 0, 0, 0, 0, 0, 0},
186 {4, 4, 2, 2, 1, 1, 1},
187 {4, 4, 2, 2 , 1, 1, 1},
188 {4, 4, 2, 2, 1, 1, 1}
191 size_t vector_length = vector_lengths[channels - 1][mat_dst[0].depth()];
193 int max_offset_cols = 0;
194 for(int i = 0; i < channels; i++)
196 int offset_cols = (mat_dst[i].offset / mat_dst[i].elemSize()) & (vector_length - 1);
197 if(max_offset_cols < offset_cols)
198 max_offset_cols = offset_cols;
201 int cols = vector_length == 1 ? divUp(mat_src.cols, vector_length)
202 : divUp(mat_src.cols + max_offset_cols, vector_length);
204 size_t localThreads[3] = { 64, 4, 1 };
205 size_t globalThreads[3] = { cols, mat_src.rows, 1 };
207 int dst_step1 = mat_dst[0].cols * mat_dst[0].elemSize();
208 vector<pair<size_t , const void *> > args;
209 args.push_back( make_pair( sizeof(cl_mem), (void *)&mat_src.data));
210 args.push_back( make_pair( sizeof(cl_int), (void *)&mat_src.step));
211 args.push_back( make_pair( sizeof(cl_int), (void *)&mat_src.offset));
212 args.push_back( make_pair( sizeof(cl_mem), (void *)&mat_dst[0].data));
213 args.push_back( make_pair( sizeof(cl_int), (void *)&mat_dst[0].step));
214 args.push_back( make_pair( sizeof(cl_int), (void *)&mat_dst[0].offset));
215 args.push_back( make_pair( sizeof(cl_mem), (void *)&mat_dst[1].data));
216 args.push_back( make_pair( sizeof(cl_int), (void *)&mat_dst[1].step));
217 args.push_back( make_pair( sizeof(cl_int), (void *)&mat_dst[1].offset));
221 args.push_back( make_pair( sizeof(cl_mem), (void *)&mat_dst[2].data));
222 args.push_back( make_pair( sizeof(cl_int), (void *)&mat_dst[2].step));
223 args.push_back( make_pair( sizeof(cl_int), (void *)&mat_dst[2].offset));
227 args.push_back( make_pair( sizeof(cl_mem), (void *)&mat_dst[3].data));
228 args.push_back( make_pair( sizeof(cl_int), (void *)&mat_dst[3].step));
229 args.push_back( make_pair( sizeof(cl_int), (void *)&mat_dst[3].offset));
232 args.push_back( make_pair( sizeof(cl_int), (void *)&mat_src.rows));
233 args.push_back( make_pair( sizeof(cl_int), (void *)&cols));
234 args.push_back( make_pair( sizeof(cl_int), (void *)&dst_step1));
236 openCLExecuteKernel(clCxt, &split_mat, kernelName, globalThreads, localThreads, args, channels, depth);
238 static void split(const oclMat &mat_src, oclMat *mat_dst)
242 int depth = mat_src.depth();
243 int num_channels = mat_src.oclchannels();
244 Size size = mat_src.size();
246 if(num_channels == 1)
248 mat_src.copyTo(mat_dst[0]);
253 for(i = 0; i < num_channels; i++)
254 mat_dst[i].create(size, CV_MAKETYPE(depth, 1));
256 split_vector_run(mat_src, mat_dst);
262 void cv::ocl::merge(const oclMat *src, size_t n, oclMat &dst)
264 split_merge::merge(src, n, dst);
266 void cv::ocl::merge(const vector<oclMat> &src, oclMat &dst)
268 split_merge::merge(&src[0], src.size(), dst);
271 void cv::ocl::split(const oclMat &src, oclMat *dst)
273 split_merge::split(src, dst);
275 void cv::ocl::split(const oclMat &src, vector<oclMat> &dst)
277 dst.resize(src.oclchannels());
278 if(src.oclchannels() > 0)
279 split_merge::split(src, &dst[0]);