added OpenCL version of cv::scaleAdd

modules/core/src/matmul.cpp

@@ -41,6 +41,7 @@
 //M*/
 
 #include "precomp.hpp"
+#include "opencl_kernels.hpp"
 #include "opencv2/core/opencl/runtime/opencl_clamdblas.hpp"
 
 #ifdef HAVE_IPP
@@ -2154,20 +2155,61 @@ static void scaleAdd_64f(const double* src1, const double* src2, double* dst,
 
 typedef void (*ScaleAddFunc)(const uchar* src1, const uchar* src2, uchar* dst, int len, const void* alpha);
 
+static bool ocl_scaleAdd( InputArray _src1, double alpha, InputArray _src2, OutputArray _dst, int type )
+{
+    int depth = CV_MAT_DEPTH(type), cn = CV_MAT_CN(type), wdepth = std::max(depth, CV_32F);
+    bool doubleSupport = ocl::Device::getDefault().doubleFPConfig() > 0;
+    Size size = _src1.size();
+
+    if ( (!doubleSupport && depth == CV_64F) || size != _src2.size() )
+        return false;
+
+    char cvt[2][50];
+    ocl::Kernel k("KF", ocl::core::arithm_oclsrc,
+                  format("-D OP_SCALE_ADD -D BINARY_OP -D dstT=%s -D workT=%s -D convertToWT1=%s"
+                         " -D srcT1=dstT -D srcT2=dstT -D convertToDT=%s%s", ocl::typeToStr(depth),
+                         ocl::typeToStr(wdepth), ocl::convertTypeStr(depth, wdepth, 1, cvt[0]),
+                         ocl::convertTypeStr(wdepth, depth, 1, cvt[1]),
+                         doubleSupport ? " -D DOUBLE_SUPPORT" : ""));
+    if (k.empty())
+        return false;
+
+    _dst.create(size, type);
+    UMat src1 = _src1.getUMat(), src2 = _src2.getUMat(), dst = _dst.getUMat();
+
+    ocl::KernelArg src1arg = ocl::KernelArg::ReadOnlyNoSize(src1),
+            src2arg = ocl::KernelArg::ReadOnlyNoSize(src2),
+            dstarg = ocl::KernelArg::WriteOnly(dst, cn);
+
+    if (wdepth == CV_32F)
+        k.args(src1arg, src2arg, dstarg, (float)alpha);
+    else
+        k.args(src1arg, src2arg, dstarg, alpha);
+
+    size_t globalsize[2] = { dst.cols * cn, dst.rows };
+    return k.run(2, globalsize, NULL, false);
+}
+
 }
 
 void cv::scaleAdd( InputArray _src1, double alpha, InputArray _src2, OutputArray _dst )
 {
-    Mat src1 = _src1.getMat(), src2 = _src2.getMat();
-    int depth = src1.depth(), cn = src1.channels();
+    int type = _src1.type(), depth = CV_MAT_DEPTH(type), cn = CV_MAT_CN(type);
+    CV_Assert( type == _src2.type() );
+
+    if (ocl::useOpenCL() && _src1.dims() <= 2 && _src2.dims() <= 2 && _dst.isUMat() &&
+            ocl_scaleAdd(_src1, alpha, _src2, _dst, type))
+        return;
 
-    CV_Assert( src1.type() == src2.type() );
     if( depth < CV_32F )
     {
         addWeighted(_src1, alpha, _src2, 1, 0, _dst, depth);
         return;
     }
 
+    Mat src1 = _src1.getMat(), src2 = _src2.getMat();
+    CV_Assert(src1.size == src2.size);
+
     _dst.create(src1.dims, src1.size, src1.type());
     Mat dst = _dst.getMat();
 

modules/core/src/opencl/arithm.cl

@@ -91,6 +91,9 @@
 
 #else
 
+    #ifndef convertToWT2
+    #define convertToWT2 convertToWT1
+    #endif
     #define srcelem1 convertToWT1(*(__global srcT1*)(srcptr1 + src1_index))
     #define srcelem2 convertToWT2(*(__global srcT2*)(srcptr2 + src2_index))
 
@@ -230,6 +233,11 @@ dstelem = v > (dstT)(0) ? log(v) : log(-v)
     workT value = srcelem1 * alpha + beta; \
     dstelem = convertToDT(value >= 0 ? value : -value)
 
+#elif defined OP_SCALE_ADD
+#undef EXTRA_PARAMS
+#define EXTRA_PARAMS , workT alpha
+#define PROCESS_ELEM dstelem = convertToDT(srcelem1 * alpha + srcelem2)
+
 #elif defined OP_CTP_AD || defined OP_CTP_AR
 #ifdef OP_CTP_AD
 #define TO_DEGREE cartToPolar *= (180 / CV_PI);

modules/core/test/ocl/test_arithm.cpp

@@ -1323,7 +1323,6 @@ OCL_TEST_P(InRange, Scalar)
     }
 }
 
-
 //////////////////////////////// ConvertScaleAbs ////////////////////////////////////////////////
 
 typedef ArithmTestBase ConvertScaleAbs;
@@ -1341,6 +1340,23 @@ OCL_TEST_P(ConvertScaleAbs, Mat)
     }
 }
 
+//////////////////////////////// ScaleAdd ////////////////////////////////////////////////
+
+typedef ArithmTestBase ScaleAdd;
+
+OCL_TEST_P(ScaleAdd, Mat)
+{
+    for (int j = 0; j < test_loop_times; j++)
+    {
+        generateTestData();
+
+        OCL_OFF(cv::scaleAdd(src1_roi, val[0], src2_roi, dst1_roi));
+        OCL_ON(cv::scaleAdd(usrc1_roi, val[0], usrc2_roi, udst1_roi));
+
+        Near(depth <= CV_32S ? 1 : 1e-6);
+    }
+}
+
 //////////////////////////////////////// Instantiation /////////////////////////////////////////
 
 OCL_INSTANTIATE_TEST_CASE_P(Arithm, Lut, Combine(::testing::Values(CV_8U, CV_8S), OCL_ALL_DEPTHS, OCL_ALL_CHANNELS, Bool(), Bool()));
@@ -1378,6 +1394,7 @@ OCL_INSTANTIATE_TEST_CASE_P(Arithm, Sqrt, Combine(::testing::Values(CV_32F, CV_6
 OCL_INSTANTIATE_TEST_CASE_P(Arithm, Normalize, Combine(OCL_ALL_DEPTHS, Values(Channels(1)), Bool()));
 OCL_INSTANTIATE_TEST_CASE_P(Arithm, InRange, Combine(OCL_ALL_DEPTHS, OCL_ALL_CHANNELS, Bool(), Bool()));
 OCL_INSTANTIATE_TEST_CASE_P(Arithm, ConvertScaleAbs, Combine(OCL_ALL_DEPTHS, OCL_ALL_CHANNELS, Bool()));
+OCL_INSTANTIATE_TEST_CASE_P(Arithm, ScaleAdd, Combine(OCL_ALL_DEPTHS, OCL_ALL_CHANNELS, Bool()));
 
 } } // namespace cvtest::ocl