remove intel guard since the code is 2 times faster on AMD too

modules/ocl/perf/perf_filters.cpp

@@ -262,13 +262,13 @@ OCL_PERF_TEST_P(SobelFixture, Sobel,
 
         oclDst.download(dst);
 
-        SANITY_CHECK(dst);
+        SANITY_CHECK(dst, 1e-3);
     }
     else if (RUN_PLAIN_IMPL)
     {
         TEST_CYCLE() cv::Sobel(src, dst, -1, dx, dy);
 
-        SANITY_CHECK(dst);
+        SANITY_CHECK(dst, 1e-3);
     }
     else
         OCL_PERF_ELSE
@@ -326,7 +326,7 @@ OCL_PERF_TEST_P(GaussianBlurFixture, GaussianBlur,
     Mat src(srcSize, type), dst(srcSize, type);
     declare.in(src, WARMUP_RNG).out(dst);
 
-    const double eps = src.depth() == CV_8U ? 1 + DBL_EPSILON : 3e-4;
+    const double eps = src.depth() == CV_8U ? 1 + DBL_EPSILON : 5e-4;
 
     if (RUN_OCL_IMPL)
     {

modules/ocl/src/filtering.cpp

@@ -774,12 +774,12 @@ static void sepFilter2D_SinglePass(const oclMat &src, oclMat &dst,
 
     option += " -D KERNEL_MATRIX_X=";
     for(int i=0; i<row_kernel.rows; i++)
-        option += cv::format("0x%x,", *reinterpret_cast<const unsigned int*>( &row_kernel.at<float>(i) ) );
+        option += cv::format("DIG(0x%x)", *reinterpret_cast<const unsigned int*>( &row_kernel.at<float>(i) ) );
     option += "0x0";
 
     option += " -D KERNEL_MATRIX_Y=";
     for(int i=0; i<col_kernel.rows; i++)
-        option += cv::format("0x%x,", *reinterpret_cast<const unsigned int*>( &col_kernel.at<float>(i) ) );
+        option += cv::format("DIG(0x%x)", *reinterpret_cast<const unsigned int*>( &col_kernel.at<float>(i) ) );
     option += "0x0";
 
     switch(src.type())
@@ -1410,7 +1410,7 @@ Ptr<FilterEngine_GPU> cv::ocl::createSeparableLinearFilter_GPU(int srcType, int
     //if image size is non-degenerate and large enough
     //and if filter support is reasonable to satisfy larger local memory requirements,
     //then we can use single pass routine to avoid extra runtime calls overhead
-    if( clCxt && clCxt->supportsFeature(FEATURE_CL_INTEL_DEVICE) &&
+    if( clCxt &&
         rowKernel.rows <= 21 && columnKernel.rows <= 21 &&
         (rowKernel.rows & 1) == 1 && (columnKernel.rows & 1) == 1 &&
         imgSize.width > optimizedSepFilterLocalSize + (rowKernel.rows>>1) &&

modules/ocl/src/opencl/filtering_sep_filter_singlepass.cl

@@ -84,6 +84,8 @@
 
 #define DST(_x,_y) (((global DSTTYPE*)(Dst+DstOffset+(_y)*DstPitch))[_x])
 
+#define DIG(a) a,
+
 //horizontal and vertical filter kernels
 //should be defined on host during compile time to avoid overhead
 __constant uint mat_kernelX[] = {KERNEL_MATRIX_X};