Merge pull request #528 from bitwangyaoyao:2.4_fixPyrLK

modules/ocl/src/mcwutil.cpp

@@ -46,6 +46,9 @@
 #include "mcwutil.hpp"
 
 #if defined (HAVE_OPENCL)
+#ifndef CL_VERSION_1_2
+#define CL_VERSION_1_2 0
+#endif
 
 using namespace std;
 
@@ -123,6 +126,101 @@ namespace cv
             openCLExecuteKernel_2(clCxt, source, kernelName, globalThreads, localThreads, args, channels, depth,
                                   build_options, finish_mode);
         }
+ 
+       cl_mem bindTexture(const oclMat &mat)
+        {
+            cl_mem texture;
+            cl_image_format format;
+            int err;
+            int depth    = mat.depth();
+            int channels = mat.channels();
+
+            switch(depth)
+            {
+            case CV_8U:
+                format.image_channel_data_type = CL_UNSIGNED_INT8;
+                break;
+            case CV_32S:
+                format.image_channel_data_type = CL_UNSIGNED_INT32;
+                break;
+            case CV_32F:
+                format.image_channel_data_type = CL_FLOAT;
+                break;
+            default:
+                throw std::exception();
+                break;
+            }
+            switch(channels)
+            {
+            case 1:
+                format.image_channel_order     = CL_R;
+                break;
+            case 3:
+                format.image_channel_order     = CL_RGB;
+                break;
+            case 4:
+                format.image_channel_order     = CL_RGBA;
+                break;
+            default:
+                throw std::exception();
+                break;
+            }
+#if CL_VERSION_1_2
+            cl_image_desc desc;
+            desc.image_type       = CL_MEM_OBJECT_IMAGE2D;
+            desc.image_width      = mat.cols;
+            desc.image_height     = mat.rows;
+            desc.image_depth      = 0;
+            desc.image_array_size = 1;
+            desc.image_row_pitch  = 0;
+            desc.image_slice_pitch = 0;
+            desc.buffer           = NULL;
+            desc.num_mip_levels   = 0;
+            desc.num_samples      = 0;
+            texture = clCreateImage(mat.clCxt->impl->clContext, CL_MEM_READ_WRITE, &format, &desc, NULL, &err);
+#else
+            texture = clCreateImage2D(
+                mat.clCxt->impl->clContext,
+                CL_MEM_READ_WRITE,
+                &format,
+                mat.cols,
+                mat.rows,
+                0,
+                NULL,
+                &err);
+#endif
+            size_t origin[] = { 0, 0, 0 };
+            size_t region[] = { mat.cols, mat.rows, 1 };
+
+            cl_mem devData;
+            if (mat.cols * mat.elemSize() != mat.step)
+            {
+                devData = clCreateBuffer(mat.clCxt->impl->clContext, CL_MEM_READ_ONLY, mat.cols * mat.rows
+                    * mat.elemSize(), NULL, NULL);
+                const size_t regin[3] = {mat.cols * mat.elemSize(), mat.rows, 1};
+                clEnqueueCopyBufferRect(mat.clCxt->impl->clCmdQueue, (cl_mem)mat.data, devData, origin, origin, 
+                    regin, mat.step, 0, mat.cols * mat.elemSize(), 0, 0, NULL, NULL);
+            }
+            else
+            {
+                devData = (cl_mem)mat.data;
+            }
+
+            clEnqueueCopyBufferToImage(mat.clCxt->impl->clCmdQueue, devData, texture, 0, origin, region, 0, NULL, 0);
+            if ((mat.cols * mat.elemSize() != mat.step))
+            {
+                clFinish(mat.clCxt->impl->clCmdQueue);
+                clReleaseMemObject(devData);
+            }
+
+            openCLSafeCall(err);
+            return texture;
+        }
+
+        void releaseTexture(cl_mem& texture)
+        {
+            openCLFree(texture);
+        }
     }//namespace ocl
 
 }//namespace cv

modules/ocl/src/mcwutil.hpp

@@ -67,6 +67,12 @@ namespace cv
         void openCLExecuteKernel2(Context *clCxt , const char **source, string kernelName, size_t globalThreads[3],
                                   size_t localThreads[3],  vector< pair<size_t, const void *> > &args, int channels,
                                   int depth, char *build_options, FLUSH_MODE finish_mode = DISABLE);
+        // bind oclMat to OpenCL image textures
+        // note:
+        //   1. there is no memory management. User need to explicitly release the resource
+        //   2. for faster clamping, there is no buffer padding for the constructed texture
+        cl_mem bindTexture(const oclMat &mat);
+        void releaseTexture(cl_mem& texture);
     }//namespace ocl
 
 }//namespace cv