added 3-channels support to cv::filter2D, cv::Laplacian

modules/imgproc/src/filter.cpp

@@ -3157,10 +3157,9 @@ static bool ocl_filter2D( InputArray _src, OutputArray _dst, int ddepth,
     if (abs(delta) > FLT_MIN)
         return false;
 
-    int type = _src.type();
-    int cn = CV_MAT_CN(type);
-    if ((1 != cn) && (2 != cn) && (4 != cn))
-        return false;//TODO
+    int type = _src.type(), cn = CV_MAT_CN(type);
+    if (cn > 4)
+        return false;
 
     int sdepth = CV_MAT_DEPTH(type);
     Size ksize = _kernel.size();
@@ -3298,7 +3297,7 @@ static bool ocl_filter2D( InputArray _src, OutputArray _dst, int ddepth,
     double borderValueDouble[4] = {0, 0, 0, 0};
     if ((borderType & ~BORDER_ISOLATED) == BORDER_CONSTANT)
     {
-        int cnocl = (3 == cn) ? 4 : cn;
+        int cnocl = 3 == cn ? 4 : cn;
         if (useDouble)
             idxArg = kernel.set(idxArg, (void *)&borderValueDouble[0], sizeof(double) * cnocl);
         else

modules/imgproc/src/opencl/filter2D.cl

@@ -203,10 +203,24 @@
 #define VEC_TYPE CAT(BASE_TYPE, VEC_SIZE)
 #define TYPE VEC_TYPE
 
+#if VEC_SIZE == 3
+#define SCALAR_TYPE CAT(FPTYPE, 4)
+#else
 #define SCALAR_TYPE CAT(FPTYPE, VEC_SIZE)
+#endif
 
 #define INTERMEDIATE_TYPE CAT(FPTYPE, VEC_SIZE)
 
+#if DATA_CHAN != 3
+#define loadpix(addr) *(__global const TYPE *)(addr)
+#define storepix(val, addr)  *(__global TYPE *)(addr) = val
+#define TSIZE (int)sizeof(TYPE)
+#else
+#define loadpix(addr) vload3(0, (__global const BASE_TYPE *)(addr))
+#define storepix(val, addr) vstore3(val, 0, (__global BASE_TYPE *)(addr))
+#define TSIZE (int)sizeof(BASE_TYPE)*DATA_CHAN
+#endif
+
 struct RectCoords
 {
     int x1, y1, x2, y2;
@@ -235,13 +249,17 @@ inline INTERMEDIATE_TYPE readSrcPixel(int2 pos, __global const uchar* srcptr, in
 #endif
     {
         //__global TYPE* ptr = (__global TYPE*)((__global char*)src + pos.x * sizeof(TYPE) + pos.y * srcStepBytes);
-        __global TYPE* ptr = (__global TYPE*)(srcptr + pos.y * srcstep + pos.x * sizeof(TYPE));
-        return CONVERT_TO_FPTYPE(*ptr);
+        __global TYPE* ptr = (__global TYPE*)(srcptr + pos.y * srcstep + pos.x * TSIZE);
+        return CONVERT_TO_FPTYPE(loadpix(ptr));
     }
     else
     {
 #ifdef BORDER_CONSTANT
+#if VEC_SIZE == 3
+        return (INTERMEDIATE_TYPE)(borderValue.x, borderValue.y, borderValue.z);
+#else
         return borderValue;
+#endif
 #else
         int selected_col = pos.x;
         int selected_row = pos.y;
@@ -262,8 +280,8 @@ inline INTERMEDIATE_TYPE readSrcPixel(int2 pos, __global const uchar* srcptr, in
         if(pos.x >= 0 && pos.y >= 0 && pos.x < srcCoords.x2 && pos.y < srcCoords.y2)
         {
             //__global TYPE* ptr = (__global TYPE*)((__global char*)src + pos.x * sizeof(TYPE) + pos.y * srcStepBytes);
-            __global TYPE* ptr = (__global TYPE*)(srcptr + pos.y * srcstep + pos.x * sizeof(TYPE));
-            return CONVERT_TO_FPTYPE(*ptr);
+            __global TYPE* ptr = (__global TYPE*)(srcptr + pos.y * srcstep + pos.x * TSIZE);
+            return CONVERT_TO_FPTYPE(loadpix(ptr));
         }
         else
         {
@@ -300,7 +318,7 @@ void filter2D(__global const uchar* srcptr, int srcstep, int srcOffsetX, int src
     int2 srcPos = (int2)(srcCoords.x1 + x, srcCoords.y1 + y - ANCHOR_Y);
 
     int2 pos = (int2)(x, y);
-    __global TYPE* dstPtr = (__global TYPE*)((__global char*)dstptr + pos.y * dststep + dstoffset + pos.x * sizeof(TYPE)); // Pointer can be out of bounds!
+    __global TYPE* dstPtr = (__global TYPE*)((__global char*)dstptr + pos.y * dststep + dstoffset + pos.x * TSIZE); // Pointer can be out of bounds!
     bool writeResult = ((local_id >= ANCHOR_X) && (local_id < LOCAL_SIZE - (KERNEL_SIZE_X - 1 - ANCHOR_X)) &&
                         (pos.x >= 0) && (pos.x < cols));
 
@@ -360,7 +378,7 @@ void filter2D(__global const uchar* srcptr, int srcstep, int srcOffsetX, int src
 
         if (writeResult)
         {
-            *dstPtr = CONVERT_TO_TYPE(total_sum);
+            storepix(CONVERT_TO_TYPE(total_sum), dstPtr);
         }
 
 #if BLOCK_SIZE_Y > 1

modules/imgproc/src/smooth.cpp

@@ -41,7 +41,6 @@
 //M*/
 
 #include "precomp.hpp"
-#define CV_OPENCL_RUN_ASSERT
 #include "opencl_kernels.hpp"
 
 /*
@@ -642,10 +641,7 @@ static bool ocl_boxFilter( InputArray _src, OutputArray _dst, int ddepth,
 
     if (cn > 4 || (!doubleSupport && (sdepth == CV_64F || ddepth == CV_64F)) ||
         _src.offset() % esz != 0 || _src.step() % esz != 0)
-    {
-        printf("!!!!!!!!!!!!!!!!!!!!!!!\n");
         return false;
-    }
 
     if (anchor.x < 0)
         anchor.x = ksize.width / 2;

modules/imgproc/test/ocl/test_filter2d.cpp

@@ -115,11 +115,10 @@ OCL_TEST_P(Filter2D, Mat)
     }
 }
 
-
 OCL_INSTANTIATE_TEST_CASE_P(ImageProc, Filter2D,
                             Combine(
-                                Values(CV_8U, CV_16U, CV_16S, CV_32F, CV_64F),
-                                Values(1, 2, 4),
+                                Values(CV_8U, CV_16U, CV_32F),
+                                OCL_ALL_CHANNELS,
                                 Values((BorderType)BORDER_CONSTANT,
                                        (BorderType)BORDER_REPLICATE,
                                        (BorderType)BORDER_REFLECT,

modules/imgproc/test/ocl/test_filters.cpp

@@ -304,7 +304,7 @@ OCL_TEST_P(MorphologyEx, Mat)
             (int)BORDER_REFLECT|BORDER_ISOLATED, (int)BORDER_WRAP|BORDER_ISOLATED, \
             (int)BORDER_REFLECT_101|BORDER_ISOLATED*/) // WRAP and ISOLATED are not supported by cv:: version
 
-#define FILTER_TYPES Values(CV_8UC1, CV_8UC3, CV_8UC4, CV_32FC1, CV_32FC3, CV_32FC4)
+#define FILTER_TYPES Values(CV_8UC1, CV_8UC3, CV_8UC4, CV_16UC1, CV_16UC3, CV_16UC4, CV_32FC1, CV_32FC3, CV_32FC4)
 
 OCL_INSTANTIATE_TEST_CASE_P(Filter, Bilateral, Combine(
                             Values(CV_8UC1, CV_8UC3),