used new stratehy in cv::accumulate**

diff --git a/modules/core/include/opencv2/core/ocl.hpp b/modules/core/include/opencv2/core/ocl.hpp
index 8a41f1b..744cf21 100644 (file)
--- a/modules/core/include/opencv2/core/ocl.hpp
+++ b/modules/core/include/opencv2/core/ocl.hpp
@@ -617,6 +617,11 @@ CV_EXPORTS int predictOptimalVectorWidth(InputArray src1, InputArray src2 = noAr
                                          InputArray src7 = noArray(), InputArray src8 = noArray(), InputArray src9 = noArray(),
                                          OclVectorStrategy strat = OCL_VECTOR_DEFAULT);
 
+// with OCL_VECTOR_MAX strategy
+CV_EXPORTS int predictOptimalVectorWidthMax(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());
+
 CV_EXPORTS void buildOptionsAddMatrixDescription(String& buildOptions, const String& name, InputArray _m);
 
 class CV_EXPORTS Image2D

diff --git a/modules/core/src/ocl.cpp b/modules/core/src/ocl.cpp
index cdc8917..0b8e033 100644 (file)
--- a/modules/core/src/ocl.cpp
+++ b/modules/core/src/ocl.cpp
@@ -4462,6 +4462,7 @@ String kernelToStr(InputArray _kernel, int ddepth, const char * name)
             offsets.push_back(src.offset()); \
             steps.push_back(src.step()); \
             dividers.push_back(ckercn * CV_ELEM_SIZE1(ctype)); \
+            kercns.push_back(ckercn); \
         } \
     } \
     while ((void)0, 0)
@@ -4483,13 +4484,13 @@ int predictOptimalVectorWidth(InputArray src1, InputArray src2, InputArray src3,
     if (vectorWidths[0] == 1)
     {
         // it's heuristic
-        vectorWidths[0] = vectorWidths[1] = 4;
-        vectorWidths[2] = vectorWidths[3] = 2;
-        vectorWidths[4] = vectorWidths[5] = vectorWidths[6] = 4;
+        vectorWidths[CV_8U] = vectorWidths[CV_8S] = 16;
+        vectorWidths[CV_16U] = vectorWidths[CV_16S] = 8;
+        vectorWidths[CV_32S] = vectorWidths[CV_32F] = vectorWidths[CV_64F] = 1;
     }
 
     std::vector<size_t> offsets, steps, cols;
-    std::vector<int> dividers;
+    std::vector<int> dividers, kercns;
     PROCESS_SRC(src1);
     PROCESS_SRC(src2);
     PROCESS_SRC(src3);
@@ -4503,23 +4504,22 @@ int predictOptimalVectorWidth(InputArray src1, InputArray src2, InputArray src3,
     size_t size = offsets.size();
 
     for (size_t i = 0; i < size; ++i)
-        while (offsets[i] % dividers[i] != 0 || steps[i] % dividers[i] != 0 || cols[i] % dividers[i] != 0)
-            dividers[i] >>= 1;
+        while (offsets[i] % dividers[i] != 0 || steps[i] % dividers[i] != 0 || cols[i] % kercns[i] != 0)
+            dividers[i] >>= 1, kercns[i] >>= 1;
 
     // default strategy
-    int kercn = *std::min_element(dividers.begin(), dividers.end());
-
-    // another strategy
-    // for (size_t i = 0; i < size; ++i)
-    //     if (dividers[i] != wsz)
-    //     {
-    //         kercn = 1;
-    //         break;
-    //     }
+    int kercn = *std::min_element(kercns.begin(), kercns.end());
 
     return kercn;
 }
 
+int predictOptimalVectorWidthMax(InputArray src1, InputArray src2, InputArray src3,
+                                 InputArray src4, InputArray src5, InputArray src6,
+                                 InputArray src7, InputArray src8, InputArray src9)
+{
+    return predictOptimalVectorWidth(src1, src2, src3, src4, src5, src6, src7, src8, src9, OCL_VECTOR_MAX);
+}
+
 #undef PROCESS_SRC
 
 

diff --git a/modules/imgproc/src/accum.cpp b/modules/imgproc/src/accum.cpp
index 3987405..9f23d34 100644 (file)
--- a/modules/imgproc/src/accum.cpp
+++ b/modules/imgproc/src/accum.cpp
@@ -370,16 +370,9 @@ static bool ocl_accumulate( InputArray _src, InputArray _src2, InputOutputArray
               op_type == ACCUMULATE_PRODUCT || op_type == ACCUMULATE_WEIGHTED);
 
     const ocl::Device & dev = ocl::Device::getDefault();
-    int vectorWidths[] = { 4, 4, 2, 2, 1, 1, 1, -1 };
+    bool haveMask = !_mask.empty(), doubleSupport = dev.doubleFPConfig() > 0;
     int stype = _src.type(), sdepth = CV_MAT_DEPTH(stype), cn = CV_MAT_CN(stype), ddepth = _dst.depth();
-    int pcn = std::max(vectorWidths[sdepth], vectorWidths[ddepth]), sesz = CV_ELEM_SIZE(sdepth) * pcn,
-        desz = CV_ELEM_SIZE(ddepth) * pcn, rowsPerWI = dev.isIntel() ? 4 : 1;
-
-    bool doubleSupport = dev.doubleFPConfig() > 0, haveMask = !_mask.empty(),
-        usepcn = _src.offset() % sesz == 0 && _src.step() % sesz == 0 && (_src.cols() * cn) % pcn == 0 &&
-            _src2.offset() % desz == 0 && _src2.step() % desz == 0 &&
-            _dst.offset() % pcn == 0 && _dst.step() % desz == 0 && !haveMask;
-    int kercn = usepcn ? pcn : haveMask ? cn : 1;
+    int kercn = haveMask ? cn : ocl::predictOptimalVectorWidthMax(_src, _src2, _dst), rowsPerWI = dev.isIntel() ? 4 : 1;
 
     if (!doubleSupport && (sdepth == CV_64F || ddepth == CV_64F))
         return false;