Merge pull request #1692 from ilya-lavrenov:ocl_separableFilter

modules/ocl/src/filtering.cpp

@@ -1058,74 +1058,39 @@ template <> struct index_and_sizeof<float>
 template <typename T>
 void linearRowFilter_gpu(const oclMat &src, const oclMat &dst, oclMat mat_kernel, int ksize, int anchor, int bordertype)
 {
-    Context *clCxt = src.clCxt;
+    CV_Assert(bordertype <= BORDER_REFLECT_101);
+    CV_Assert(ksize == (anchor << 1) + 1);
     int channels = src.oclchannels();
 
-    size_t localThreads[3] = {16, 16, 1};
-    string kernelName = "row_filter";
-
-    char btype[30];
+    size_t localThreads[3] = { 16, 16, 1 };
+    size_t globalThreads[3] = { dst.cols, dst.rows, 1 };
 
-    switch (bordertype)
-    {
-    case 0:
-        sprintf(btype, "BORDER_CONSTANT");
-        break;
-    case 1:
-        sprintf(btype, "BORDER_REPLICATE");
-        break;
-    case 2:
-        sprintf(btype, "BORDER_REFLECT");
-        break;
-    case 3:
-        sprintf(btype, "BORDER_WRAP");
-        break;
-    case 4:
-        sprintf(btype, "BORDER_REFLECT_101");
-        break;
-    }
-
-    char compile_option[128];
-    sprintf(compile_option, "-D RADIUSX=%d -D LSIZE0=%d -D LSIZE1=%d -D CN=%d -D %s", anchor, (int)localThreads[0], (int)localThreads[1], channels, btype);
-
-    size_t globalThreads[3];
-    globalThreads[1] = (dst.rows + localThreads[1] - 1) / localThreads[1] * localThreads[1];
-    globalThreads[2] = (1 + localThreads[2] - 1) / localThreads[2] * localThreads[2];
+    const char * const borderMap[] = { "BORDER_CONSTANT", "BORDER_REPLICATE", "BORDER_REFLECT", "BORDER_WRAP", "BORDER_REFLECT_101" };
+    std::string buildOptions = format("-D RADIUSX=%d -D LSIZE0=%d -D LSIZE1=%d -D CN=%d -D %s",
+            anchor, (int)localThreads[0], (int)localThreads[1], channels, borderMap[bordertype]);
 
     if (src.depth() == CV_8U)
     {
         switch (channels)
         {
         case 1:
-        case 3:
-            globalThreads[0] = ((dst.cols + 4) / 4 + localThreads[0] - 1) / localThreads[0] * localThreads[0];
+            globalThreads[0] = (dst.cols + 3) >> 2;
             break;
         case 2:
-            globalThreads[0] = ((dst.cols + 1) / 2 + localThreads[0] - 1) / localThreads[0] * localThreads[0];
+            globalThreads[0] = (dst.cols + 1) >> 1;
             break;
         case 4:
-            globalThreads[0] = (dst.cols + localThreads[0] - 1) / localThreads[0] * localThreads[0];
+            globalThreads[0] = dst.cols;
             break;
         }
     }
-    else
-    {
-        globalThreads[0] = (dst.cols + localThreads[0] - 1) / localThreads[0] * localThreads[0];
-    }
 
-    //sanity checks
-    CV_Assert(clCxt == dst.clCxt);
-    CV_Assert(src.cols == dst.cols);
-    CV_Assert(src.oclchannels() == dst.oclchannels());
-    CV_Assert(ksize == (anchor << 1) + 1);
-    int src_pix_per_row, dst_pix_per_row;
-    int src_offset_x, src_offset_y;//, dst_offset_in_pixel;
-    src_pix_per_row = src.step / src.elemSize();
-    src_offset_x = (src.offset % src.step) / src.elemSize();
-    src_offset_y = src.offset / src.step;
-    dst_pix_per_row = dst.step / dst.elemSize();
-    //dst_offset_in_pixel = dst.offset / dst.elemSize();
+    int src_pix_per_row = src.step / src.elemSize();
+    int src_offset_x = (src.offset % src.step) / src.elemSize();
+    int src_offset_y = src.offset / src.step;
+    int dst_pix_per_row = dst.step / dst.elemSize();
     int ridusy = (dst.rows - src.rows) >> 1;
+
     vector<pair<size_t , const void *> > args;
     args.push_back(make_pair(sizeof(cl_mem), &src.data));
     args.push_back(make_pair(sizeof(cl_mem), &dst.data));
@@ -1140,7 +1105,8 @@ void linearRowFilter_gpu(const oclMat &src, const oclMat &dst, oclMat mat_kernel
     args.push_back(make_pair(sizeof(cl_int), (void *)&ridusy));
     args.push_back(make_pair(sizeof(cl_mem), (void *)&mat_kernel.data));
 
-    openCLExecuteKernel(clCxt, &filter_sep_row, kernelName, globalThreads, localThreads, args, channels, src.depth(), compile_option);
+    openCLExecuteKernel(src.clCxt, &filter_sep_row, "row_filter", globalThreads, localThreads,
+                        args, channels, src.depth(), buildOptions.c_str());
 }
 
 Ptr<BaseRowFilter_GPU> cv::ocl::getLinearRowFilter_GPU(int srcType, int /*bufType*/, const Mat &rowKernel, int anchor, int bordertype)

modules/ocl/src/opencl/filter_sep_col.cl

@@ -47,36 +47,6 @@
 #define READ_TIMES_ROW ((2*(RADIUS+LSIZE0)-1)/LSIZE0)
 #endif
 
-#ifdef BORDER_CONSTANT
-//BORDER_CONSTANT:      iiiiii|abcdefgh|iiiiiii
-#define ELEM(i,l_edge,r_edge,elem1,elem2) (i)<(l_edge) | (i) >= (r_edge) ? (elem1) : (elem2)
-#endif
-
-#ifdef BORDER_REPLICATE
-//BORDER_REPLICATE:     aaaaaa|abcdefgh|hhhhhhh
-#define ADDR_L(i,l_edge,r_edge)  (i) < (l_edge) ? (l_edge) : (i)
-#define ADDR_R(i,r_edge,addr)   (i) >= (r_edge) ? (r_edge)-1 : (addr)
-#endif
-
-#ifdef BORDER_REFLECT
-//BORDER_REFLECT:       fedcba|abcdefgh|hgfedcb
-#define ADDR_L(i,l_edge,r_edge)  (i) < (l_edge) ? -(i)-1 : (i)
-#define ADDR_R(i,r_edge,addr) (i) >= (r_edge) ? -(i)-1+((r_edge)<<1) : (addr)
-#endif
-
-#ifdef BORDER_REFLECT_101
-//BORDER_REFLECT_101:   gfedcb|abcdefgh|gfedcba
-#define ADDR_L(i,l_edge,r_edge)  (i) < (l_edge) ? -(i) : (i)
-#define ADDR_R(i,r_edge,addr) (i) >= (r_edge) ? -(i)-2+((r_edge)<<1) : (addr)
-#endif
-
-#ifdef BORDER_WRAP
-//BORDER_WRAP:          cdefgh|abcdefgh|abcdefg
-#define ADDR_L(i,l_edge,r_edge)  (i) < (l_edge) ? (i)+(r_edge) : (i)
-#define ADDR_R(i,r_edge,addr)   (i) >= (r_edge) ?   (i)-(r_edge) : (addr)
-#endif
-
-
 /**********************************************************************************
 These kernels are written for separable filters such as Sobel, Scharr, GaussianBlur.
 Now(6/29/2011) the kernels only support 8U data type and the anchor of the convovle
@@ -107,15 +77,16 @@ __kernel __attribute__((reqd_work_group_size(LSIZE0,LSIZE1,1))) void col_filter
 {
     int x = get_global_id(0);
     int y = get_global_id(1);
+
     int l_x = get_local_id(0);
     int l_y = get_local_id(1);
-    int start_addr = mad24(y,src_step_in_pixel,x);
-    int end_addr = mad24(src_whole_rows - 1,src_step_in_pixel,src_whole_cols);
-    int i;
-    GENTYPE_SRC sum;
-    GENTYPE_SRC temp[READ_TIMES_COL];
 
-    __local GENTYPE_SRC LDS_DAT[LSIZE1*READ_TIMES_COL][LSIZE0+1];
+    int start_addr = mad24(y, src_step_in_pixel, x);
+    int end_addr = mad24(src_whole_rows - 1, src_step_in_pixel, src_whole_cols);
+
+    int i;
+    GENTYPE_SRC sum, temp[READ_TIMES_COL];
+    __local GENTYPE_SRC LDS_DAT[LSIZE1 * READ_TIMES_COL][LSIZE0 + 1];
 
     //read pixels from src
     for(i = 0;i<READ_TIMES_COL;i++)

modules/ocl/test/test_filters.cpp

@@ -403,7 +403,7 @@ INSTANTIATE_TEST_CASE_P(Filter, SobelTest, Combine(
                             Bool()));
 
 INSTANTIATE_TEST_CASE_P(Filter, ScharrTest, Combine(
-                            Values(CV_8UC1, CV_8UC3, CV_8UC4, CV_32FC1, CV_32FC4),
+                            Values(CV_8UC1, CV_8UC3, CV_8UC4, CV_32FC1, CV_32FC3, CV_32FC4),
                             Values(0), // not used
                             Values(Size(0, 1), Size(1, 0)),
                             Values((int)BORDER_CONSTANT, (int)BORDER_REFLECT101,