Merge pull request #3354 from vbystricky:oclopt_convertScaleAbs

modules/core/include/opencv2/core/ocl.hpp

@@ -618,7 +618,7 @@ CV_EXPORTS int predictOptimalVectorWidth(InputArray src1, InputArray src2 = noAr
                                          InputArray src7 = noArray(), InputArray src8 = noArray(), InputArray src9 = noArray(),
                                          OclVectorStrategy strat = OCL_VECTOR_DEFAULT);
 
-CV_EXPORTS int checkOptimalVectorWidth(int *vectorWidths,
+CV_EXPORTS int checkOptimalVectorWidth(const int *vectorWidths,
                                        InputArray src1, InputArray src2 = noArray(), InputArray src3 = noArray(),
                                        InputArray src4 = noArray(), InputArray src5 = noArray(), InputArray src6 = noArray(),
                                        InputArray src7 = noArray(), InputArray src8 = noArray(), InputArray src9 = noArray(),

modules/core/src/convert.cpp

@@ -3275,13 +3275,26 @@ static BinaryFunc getConvertScaleFunc(int sdepth, int ddepth)
 static bool ocl_convertScaleAbs( InputArray _src, OutputArray _dst, double alpha, double beta )
 {
     const ocl::Device & d = ocl::Device::getDefault();
-    int type = _src.type(), depth = CV_MAT_DEPTH(type), cn = CV_MAT_CN(type),
-        kercn = ocl::predictOptimalVectorWidth(_src, _dst), rowsPerWI = d.isIntel() ? 4 : 1;
-    bool doubleSupport = d.doubleFPConfig() > 0;
 
+    int type = _src.type(), depth = CV_MAT_DEPTH(type), cn = CV_MAT_CN(type);
+    bool doubleSupport = d.doubleFPConfig() > 0;
     if (!doubleSupport && depth == CV_64F)
         return false;
 
+    _dst.create(_src.size(), CV_8UC(cn));
+    int kercn = 1;
+    if (d.isIntel())
+    {
+        static const int vectorWidths[] = {4, 4, 4, 4, 4, 4, 4, -1};
+        kercn = ocl::checkOptimalVectorWidth( vectorWidths, _src, _dst,
+                                              noArray(), noArray(), noArray(),
+                                              noArray(), noArray(), noArray(),
+                                              noArray(), ocl::OCL_VECTOR_MAX);
+    }
+    else
+        kercn = ocl::predictOptimalVectorWidthMax(_src, _dst);
+
+    int rowsPerWI = d.isIntel() ? 4 : 1;
     char cvt[2][50];
     int wdepth = std::max(depth, CV_32F);
     String build_opt = format("-D OP_CONVERT_SCALE_ABS -D UNARY_OP -D dstT=%s -D srcT1=%s"
@@ -3299,7 +3312,6 @@ static bool ocl_convertScaleAbs( InputArray _src, OutputArray _dst, double alpha
         return false;
 
     UMat src = _src.getUMat();
-    _dst.create(src.size(), CV_8UC(cn));
     UMat dst = _dst.getUMat();
 
     ocl::KernelArg srcarg = ocl::KernelArg::ReadOnlyNoSize(src),

modules/core/src/ocl.cpp

@@ -4536,12 +4536,14 @@ int predictOptimalVectorWidth(InputArray src1, InputArray src2, InputArray src3,
     return checkOptimalVectorWidth(vectorWidths, src1, src2, src3, src4, src5, src6, src7, src8, src9, strat);
 }
 
-int checkOptimalVectorWidth(int *vectorWidths,
+int checkOptimalVectorWidth(const int *vectorWidths,
                             InputArray src1, InputArray src2, InputArray src3,
                             InputArray src4, InputArray src5, InputArray src6,
                             InputArray src7, InputArray src8, InputArray src9,
                             OclVectorStrategy strat)
 {
+    CV_Assert(vectorWidths);
+
     int ref_type = src1.type();
 
     std::vector<size_t> offsets, steps, cols;