core(ocl): options to control buffer access flags

- control using of clEnqueueMapBuffer or clEnqueueReadBuffer[Rect]
- added benchmarks with OpenCL buffer access use cases

modules/core/perf/opencl/perf_matop.cpp

@@ -146,6 +146,141 @@ OCL_PERF_TEST_P(CopyToFixture, CopyToWithMaskUninit,
     SANITY_CHECK(dst);
 }
 
+
+
+enum ROIType {
+    ROI_FULL,
+    ROI_2_RECT,
+    ROI_2_TOP,  // contiguous memory block
+    ROI_2_LEFT,
+    ROI_4,
+    ROI_16,
+};
+static Rect getROI(enum ROIType t, const Size& sz)
+{
+    switch (t)
+    {
+        case ROI_FULL: return Rect(0, 0, sz.width, sz.height);
+        case ROI_2_RECT: return Rect(0, 0, sz.width * 71 / 100, sz.height * 71 / 100);  // 71 = sqrt(1/2) * 100
+        case ROI_2_TOP: return Rect(0, 0, sz.width, sz.height / 2);  // 71 = sqrt(1/2) * 100
+        case ROI_2_LEFT: return Rect(0, 0, sz.width / 2, sz.height);  // 71 = sqrt(1/2) * 100
+        case ROI_4: return Rect(0, 0, sz.width / 2, sz.height / 2);
+        case ROI_16: return Rect(0, 0, sz.width / 4, sz.height / 4);
+    }
+    CV_Assert(false);
+}
+
+typedef TestBaseWithParam< tuple<cv::Size, MatType, ROIType> > OpenCLBuffer;
+
+static inline void PrintTo(const tuple<cv::Size, MatType, enum ROIType>& v, std::ostream* os)
+{
+    *os << "(" << get<0>(v) << ", " << typeToString(get<1>(v)) << ", ";
+    enum ROIType roiType = get<2>(v);
+    if (roiType == ROI_FULL)
+        *os << "ROI_100_FULL";
+    else if (roiType == ROI_2_RECT)
+        *os << "ROI_050_RECT_HALF";
+    else if (roiType == ROI_2_TOP)
+        *os << "ROI_050_TOP_HALF";
+    else if (roiType == ROI_2_LEFT)
+        *os << "ROI_050_LEFT_HALF";
+    else if (roiType == ROI_4)
+        *os << "ROI_025_1/4";
+    else
+        *os << "ROI_012_1/16";
+    *os << ")";
+}
+
+PERF_TEST_P_(OpenCLBuffer, cpu_write)
+{
+    const Size srcSize = get<0>(GetParam());
+    const int type = get<1>(GetParam());
+    const Rect roi = getROI(get<2>(GetParam()), srcSize);
+
+    checkDeviceMaxMemoryAllocSize(srcSize, type);
+
+    UMat src(srcSize, type);
+    declare.in(src(roi), WARMUP_NONE);
+
+    OCL_TEST_CYCLE()
+    {
+        Mat m = src(roi).getMat(ACCESS_WRITE);
+        m.setTo(Scalar(1, 2, 3, 4));
+    }
+
+    SANITY_CHECK_NOTHING();
+}
+
+PERF_TEST_P_(OpenCLBuffer, cpu_read)
+{
+    const Size srcSize = get<0>(GetParam());
+    const int type = get<1>(GetParam());
+    const Rect roi = getROI(get<2>(GetParam()), srcSize);
+
+    checkDeviceMaxMemoryAllocSize(srcSize, type);
+
+    UMat src(srcSize, type, Scalar(1, 2, 3, 4));
+    declare.in(src(roi), WARMUP_NONE);
+
+    OCL_TEST_CYCLE()
+    {
+        unsigned counter = 0;
+        Mat m = src(roi).getMat(ACCESS_READ);
+        for (int y = 0; y < m.rows; y++)
+        {
+            uchar* ptr = m.ptr(y);
+            size_t width_bytes = m.cols * m.elemSize();
+            for (size_t x_bytes = 0; x_bytes < width_bytes; x_bytes++)
+                counter += (unsigned)(ptr[x_bytes]);
+        }
+    }
+
+    SANITY_CHECK_NOTHING();
+}
+
+PERF_TEST_P_(OpenCLBuffer, cpu_update)
+{
+    const Size srcSize = get<0>(GetParam());
+    const int type = get<1>(GetParam());
+    const Rect roi = getROI(get<2>(GetParam()), srcSize);
+
+    checkDeviceMaxMemoryAllocSize(srcSize, type);
+
+    UMat src(srcSize, type, Scalar(1, 2, 3, 4));
+    declare.in(src(roi), WARMUP_NONE);
+
+    OCL_TEST_CYCLE()
+    {
+        Mat m = src(roi).getMat(ACCESS_READ | ACCESS_WRITE);
+        for (int y = 0; y < m.rows; y++)
+        {
+            uchar* ptr = m.ptr(y);
+            size_t width_bytes = m.cols * m.elemSize();
+            for (size_t x_bytes = 0; x_bytes < width_bytes; x_bytes++)
+                ptr[x_bytes] += 1;
+        }
+    }
+
+    SANITY_CHECK_NOTHING();
+}
+
+INSTANTIATE_TEST_CASE_P(/*FULL*/, OpenCLBuffer,
+    testing::Combine(
+        testing::Values(szVGA, sz720p, sz1080p, sz2160p),
+        testing::Values(CV_8UC1, CV_8UC2, CV_8UC3, CV_8UC4),
+        testing::Values(ROI_FULL)
+    )
+);
+
+INSTANTIATE_TEST_CASE_P(ROI, OpenCLBuffer,
+    testing::Combine(
+        testing::Values(sz1080p, sz2160p),
+        testing::Values(CV_8UC1),
+        testing::Values(ROI_16, ROI_4, ROI_2_RECT, ROI_2_LEFT, ROI_2_TOP, ROI_FULL)
+    )
+);
+
+
 } } // namespace opencv_test::ocl
 
 #endif // HAVE_OPENCL

modules/core/src/ocl.cpp

@@ -4607,6 +4607,17 @@ public:
         return u;
     }
 
+    static bool isOpenCLMapForced()  // force clEnqueueMapBuffer / clEnqueueUnmapMemObject OpenCL API
+    {
+        static bool value = cv::utils::getConfigurationParameterBool("OPENCV_OPENCL_BUFFER_FORCE_MAPPING", false);
+        return value;
+    }
+    static bool isOpenCLCopyingForced()  // force clEnqueueReadBuffer[Rect] / clEnqueueWriteBuffer[Rect] OpenCL API
+    {
+        static bool value = cv::utils::getConfigurationParameterBool("OPENCV_OPENCL_BUFFER_FORCE_COPYING", false);
+        return value;
+    }
+
     void getBestFlags(const Context& ctx, int /*flags*/, UMatUsageFlags usageFlags, int& createFlags, int& flags0) const
     {
         const Device& dev = ctx.device(0);
@@ -4614,7 +4625,15 @@ public:
         if ((usageFlags & USAGE_ALLOCATE_HOST_MEMORY) != 0)
             createFlags |= CL_MEM_ALLOC_HOST_PTR;
 
-        if( dev.hostUnifiedMemory() )
+        if (!isOpenCLCopyingForced() &&
+            (isOpenCLMapForced() ||
+                (dev.hostUnifiedMemory()
+#ifndef __APPLE__
+                || dev.isIntel()
+#endif
+                )
+            )
+        )
             flags0 = 0;
         else
             flags0 = UMatData::COPY_ON_MAP;