fixed bug in cv::ocl::predictOptimalVectorWidth

modules/core/src/arithm.cpp

@@ -1299,7 +1299,7 @@ static bool ocl_arithm_op(InputArray _src1, InputArray _src2, OutputArray _dst,
     int type1 = _src1.type(), depth1 = CV_MAT_DEPTH(type1), cn = CV_MAT_CN(type1);
     bool haveMask = !_mask.empty();
 
-    if( ((haveMask || haveScalar) && cn > 4) )
+    if ( (haveMask || haveScalar) && cn > 4 )
         return false;
 
     int dtype = _dst.type(), ddepth = CV_MAT_DEPTH(dtype), wdepth = std::max(CV_32S, CV_MAT_DEPTH(wtype));
@@ -1320,14 +1320,11 @@ static bool ocl_arithm_op(InputArray _src1, InputArray _src2, OutputArray _dst,
             "-D convertToWT2=%s -D convertToDT=%s%s -D cn=%d",
             (haveMask ? "MASK_" : ""), (haveScalar ? "UNARY_OP" : "BINARY_OP"),
             oclop2str[oclop], ocl::typeToStr(CV_MAKETYPE(depth1, kercn)),
-            ocl::typeToStr(CV_MAKETYPE(depth1, 1)),
-            ocl::typeToStr(CV_MAKETYPE(depth2, kercn)),
-            ocl::typeToStr(CV_MAKETYPE(depth2, 1)),
-            ocl::typeToStr(CV_MAKETYPE(ddepth, kercn)),
-            ocl::typeToStr(CV_MAKETYPE(ddepth, 1)),
-            ocl::typeToStr(CV_MAKETYPE(wdepth, kercn)),
+            ocl::typeToStr(depth1), ocl::typeToStr(CV_MAKETYPE(depth2, kercn)),
+            ocl::typeToStr(depth2), ocl::typeToStr(CV_MAKETYPE(ddepth, kercn)),
+            ocl::typeToStr(ddepth), ocl::typeToStr(CV_MAKETYPE(wdepth, kercn)),
             ocl::typeToStr(CV_MAKETYPE(wdepth, scalarcn)),
-            ocl::typeToStr(CV_MAKETYPE(wdepth, 1)), wdepth,
+            ocl::typeToStr(wdepth), wdepth,
             ocl::convertTypeStr(depth1, wdepth, kercn, cvtstr[0]),
             ocl::convertTypeStr(depth2, wdepth, kercn, cvtstr[1]),
             ocl::convertTypeStr(wdepth, ddepth, kercn, cvtstr[2]),
@@ -1347,7 +1344,7 @@ static bool ocl_arithm_op(InputArray _src1, InputArray _src2, OutputArray _dst,
     }
 
     ocl::Kernel k("KF", ocl::core::arithm_oclsrc, opts);
-    if( k.empty() )
+    if (k.empty())
         return false;
 
     UMat src1 = _src1.getUMat(), src2;
@@ -1388,12 +1385,12 @@ static bool ocl_arithm_op(InputArray _src1, InputArray _src2, OutputArray _dst,
 
         if( !haveMask )
         {
-            if(n == 0)
+            if (n == 0)
                 k.args(src1arg, src2arg, dstarg);
-            else if(n == 1)
+            else if (n == 1)
                 k.args(src1arg, src2arg, dstarg,
                        ocl::KernelArg(0, 0, 0, 0, usrdata_p, usrdata_esz));
-            else if(n == 3)
+            else if (n == 3)
                 k.args(src1arg, src2arg, dstarg,
                        ocl::KernelArg(0, 0, 0, 0, usrdata_p, usrdata_esz),
                        ocl::KernelArg(0, 0, 0, 0, usrdata_p + usrdata_esz, usrdata_esz),

modules/core/src/mathfuncs.cpp

@@ -2041,7 +2041,7 @@ static bool ocl_pow(InputArray _src, double power, OutputArray _dst,
     const char * const op = issqrt ? "OP_SQRT" : is_ipower ? "OP_POWN" : "OP_POW";
 
     ocl::Kernel k("KF", ocl::core::arithm_oclsrc,
-                  format("-D dstT=%s -D %s -D UNARY_OP%s", ocl::typeToStr(CV_MAKE_TYPE(depth, 1)),
+                  format("-D dstT=%s -D %s -D UNARY_OP%s", ocl::typeToStr(depth),
                          op, doubleSupport ? " -D DOUBLE_SUPPORT" : ""));
     if (k.empty())
         return false;
@@ -2081,7 +2081,7 @@ void pow( InputArray _src, double power, OutputArray _dst )
     {
         if( ipower < 0 )
         {
-            divide( 1., _src, _dst );
+            divide( Scalar::all(1), _src, _dst );
             if( ipower == -1 )
                 return;
             ipower = -ipower;
@@ -2115,10 +2115,7 @@ void pow( InputArray _src, double power, OutputArray _dst )
 
     Mat src, dst;
     if (same)
-    {
-        dst = _dst.getMat();
-        src = dst;
-    }
+        src = dst = _dst.getMat();
     else
     {
         src = _src.getMat();

modules/core/src/ocl.cpp

@@ -4347,7 +4347,7 @@ int predictOptimalVectorWidth(InputArray src1, InputArray src2, InputArray src3,
                               InputArray src4, InputArray src5, InputArray src6,
                               InputArray src7, InputArray src8, InputArray src9)
 {
-    int type = src1.type(), depth = CV_MAT_DEPTH(type), cn = CV_MAT_CN(type);
+    int type = src1.type(), depth = CV_MAT_DEPTH(type), cn = CV_MAT_CN(type), esz = CV_ELEM_SIZE(depth);
     Size ssize = src1.size();
     const ocl::Device & d = ocl::Device::getDefault();
 
@@ -4371,7 +4371,8 @@ int predictOptimalVectorWidth(InputArray src1, InputArray src2, InputArray src3,
     PROCESS_SRC(src9);
 
     size_t size = offsets.size();
-    std::vector<int> dividers(size, width);
+    int wsz = width * esz;
+    std::vector<int> dividers(size, wsz);
 
     for (size_t i = 0; i < size; ++i)
         while (offsets[i] % dividers[i] != 0 || steps[i] % dividers[i] != 0 || cols[i] % dividers[i] != 0)
@@ -4379,7 +4380,7 @@ int predictOptimalVectorWidth(InputArray src1, InputArray src2, InputArray src3,
 
     // default strategy
     for (size_t i = 0; i < size; ++i)
-        if (dividers[i] != width)
+        if (dividers[i] != wsz)
         {
             width = 1;
             break;

modules/core/test/ocl/test_arithm.cpp

@@ -839,7 +839,7 @@ OCL_TEST_P(Pow, Mat)
             OCL_OFF(cv::pow(src1_roi, pows[k], dst1_roi));
             OCL_ON(cv::pow(usrc1_roi, pows[k], udst1_roi));
 
-            OCL_EXPECT_MATS_NEAR_RELATIVE(dst1, 1e-6);
+            OCL_EXPECT_MATS_NEAR_RELATIVE(dst1, 1e-5);
         }
 }