Merge pull request #9882 from pengli:ocl4dnn

modules/dnn/include/opencv2/dnn/dnn.hpp

@@ -187,16 +187,26 @@ CV__DNN_EXPERIMENTAL_NS_BEGIN
          */
         virtual void forward(std::vector<Mat*> &input, std::vector<Mat> &output, std::vector<Mat> &internals) = 0;
 
+        /** @brief Given the @p input blobs, computes the output @p blobs.
+         *  @param[in]  inputs  the input blobs.
+         *  @param[out] outputs allocated output blobs, which will store results of the computation.
+         *  @param[out] internals allocated internal blobs
+         */
+        virtual void forward(InputArrayOfArrays inputs, OutputArrayOfArrays outputs, OutputArrayOfArrays internals) = 0;
+
+        /** @brief Given the @p input blobs, computes the output @p blobs.
+         *  @param[in]  inputs  the input blobs.
+         *  @param[out] outputs allocated output blobs, which will store results of the computation.
+         *  @param[out] internals allocated internal blobs
+         */
+        void forward_fallback(InputArrayOfArrays inputs, OutputArrayOfArrays outputs, OutputArrayOfArrays internals);
+
         /** @brief @overload */
         CV_WRAP void finalize(const std::vector<Mat> &inputs, CV_OUT std::vector<Mat> &outputs);
 
         /** @brief @overload */
         CV_WRAP std::vector<Mat> finalize(const std::vector<Mat> &inputs);
 
-        /** @brief @overload */
-        CV_WRAP void forward(const std::vector<Mat> &inputs, CV_IN_OUT std::vector<Mat> &outputs,
-                             CV_IN_OUT std::vector<Mat> &internals);
-
         /** @brief Allocates layer and computes output. */
         CV_WRAP void run(const std::vector<Mat> &inputs, CV_OUT std::vector<Mat> &outputs,
                          CV_IN_OUT std::vector<Mat> &internals);

modules/dnn/include/opencv2/dnn/shape_utils.hpp

@@ -132,6 +132,11 @@ static inline MatShape shape(const Mat& mat)
     return shape(mat.size.p, mat.dims);
 }
 
+static inline MatShape shape(const UMat& mat)
+{
+    return shape(mat.size.p, mat.dims);
+}
+
 namespace {inline bool is_neg(int i) { return i < 0; }}
 
 static inline MatShape shape(int a0, int a1=-1, int a2=-1, int a3=-1)
@@ -151,7 +156,7 @@ static inline int total(const MatShape& shape, int start = -1, int end = -1)
         return 0;
 
     int elems = 1;
-    CV_Assert(start < (int)shape.size() && end <= (int)shape.size() &&
+    CV_Assert(start <= (int)shape.size() && end <= (int)shape.size() &&
               start <= end);
     for(int i = start; i < end; i++)
     {

modules/dnn/src/layers/batch_norm_layer.cpp

@@ -102,6 +102,14 @@ public:
                backendId == DNN_BACKEND_HALIDE && haveHalide();
     }
 
+    void forward(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr, OutputArrayOfArrays internals_arr)
+    {
+        CV_TRACE_FUNCTION();
+        CV_TRACE_ARG_VALUE(name, "name", name.c_str());
+
+        Layer::forward_fallback(inputs_arr, outputs_arr, internals_arr);
+    }
+
     void forward(std::vector<Mat*> &inputs, std::vector<Mat> &outputs, std::vector<Mat> &internals)
     {
         CV_TRACE_FUNCTION();

modules/dnn/src/layers/blank_layer.cpp

@@ -62,6 +62,25 @@ public:
         return true;
     }
 
+#ifdef HAVE_OPENCL
+    bool forward_ocl(InputArrayOfArrays inputs, OutputArrayOfArrays outputs, OutputArrayOfArrays internals)
+    {
+        return true;
+    }
+#endif
+
+    void forward(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr, OutputArrayOfArrays internals_arr)
+    {
+        CV_TRACE_FUNCTION();
+        CV_TRACE_ARG_VALUE(name, "name", name.c_str());
+
+        CV_OCL_RUN((preferableTarget == DNN_TARGET_OPENCL) &&
+                   OCL_PERFORMANCE_CHECK(ocl::Device::getDefault().isIntel()),
+                   forward_ocl(inputs_arr, outputs_arr, internals_arr))
+
+        Layer::forward_fallback(inputs_arr, outputs_arr, internals_arr);
+    }
+
     void forward(std::vector<Mat*> &inputs, std::vector<Mat> &outputs, std::vector<Mat> &internals)
     {
         CV_TRACE_FUNCTION();

modules/dnn/src/layers/concat_layer.cpp

@@ -176,36 +176,38 @@ public:
     };
 
 #ifdef HAVE_OPENCL
-    bool forward_ocl(std::vector<Mat*> &inputs, std::vector<Mat> &outputs, std::vector<Mat> &internals)
+    bool forward_ocl(InputArrayOfArrays inps, OutputArrayOfArrays outs, OutputArrayOfArrays internals)
     {
-        CV_TRACE_FUNCTION();
-        CV_TRACE_ARG_VALUE(name, "name", name.c_str());
+        std::vector<UMat> inputs;
+        std::vector<UMat> outputs;
 
-        int cAxis = clamp(axis, inputs[0]->dims);
+        inps.getUMatVector(inputs);
+        outs.getUMatVector(outputs);
+
+        int cAxis = clamp(axis, inputs[0].dims);
         if (!(cAxis == 1 && outputs[0].dims == 4 && !padding))
             return false;
 
         int bottom_concat_axis;
-        int concat_size = inputs[0]->size[2] * inputs[0]->size[3];
+        int concat_size = inputs[0].size[2] * inputs[0].size[3];
         int top_concat_axis = outputs[0].size[1];
         int offset_concat_axis = 0;
-        UMat inpMat, outMat;
-        outMat = outputs[0].getUMat(ACCESS_WRITE);
-
-        ocl::Kernel kernel;
-        String buildopt = String("-DDtype=") + ocl::typeToStr(inputs[0]->type()) + String(" ");
-        if (!kernel.create("concat", ocl::dnn::concat_oclsrc, buildopt))
-            return false;
+        UMat& outMat = outputs[0];
+        String buildopt = String("-DDtype=") + ocl::typeToStr(inputs[0].type()) + String(" ");
 
         for (size_t i = 0; i < inputs.size(); i++)
         {
-            inpMat = inputs[i]->getUMat(ACCESS_READ);
-            bottom_concat_axis = inputs[i]->size[1];
-            size_t nthreads = inputs[i]->total();
+            ocl::Kernel kernel("concat", ocl::dnn::concat_oclsrc, buildopt);
+            if (kernel.empty())
+                return false;
+
+            UMat& inpMat = inputs[i];
+            bottom_concat_axis = inputs[i].size[1];
+            size_t nthreads = inputs[i].total();
 
             kernel.set(0, (int)nthreads);
             kernel.set(1, ocl::KernelArg::PtrReadOnly(inpMat));
-            kernel.set(2, (int)inputs[i]->size[0]);
+            kernel.set(2, (int)inputs[i].size[0]);
             kernel.set(3, (int)concat_size);
             kernel.set(4, (int)top_concat_axis);
             kernel.set(5, (int)bottom_concat_axis);
@@ -222,14 +224,22 @@ public:
     }
 #endif
 
-    void forward(std::vector<Mat*> &inputs, std::vector<Mat> &outputs, std::vector<Mat> &internals)
+    void forward(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr, OutputArrayOfArrays internals_arr)
     {
         CV_TRACE_FUNCTION();
         CV_TRACE_ARG_VALUE(name, "name", name.c_str());
 
         CV_OCL_RUN((preferableTarget == DNN_TARGET_OPENCL) &&
                    OCL_PERFORMANCE_CHECK(ocl::Device::getDefault().isIntel()),
-                   forward_ocl(inputs, outputs, internals))
+                   forward_ocl(inputs_arr, outputs_arr, internals_arr))
+
+        Layer::forward_fallback(inputs_arr, outputs_arr, internals_arr);
+    }
+
+    void forward(std::vector<Mat*> &inputs, std::vector<Mat> &outputs, std::vector<Mat> &internals)
+    {
+        CV_TRACE_FUNCTION();
+        CV_TRACE_ARG_VALUE(name, "name", name.c_str());
 
         int cAxis = clamp(axis, inputs[0]->dims);
         Mat& outMat = outputs[0];

modules/dnn/src/layers/convolution_layer.cpp

@@ -671,14 +671,20 @@ public:
     };
 
 #ifdef HAVE_OPENCL
-    bool forward_ocl(std::vector<Mat*> &inputs, std::vector<Mat> &outputs, std::vector<Mat> &internals)
+    bool forward_ocl(InputArrayOfArrays inps, OutputArrayOfArrays outs, OutputArrayOfArrays internals)
     {
-        int group = inputs[0]->size[1] / umat_blobs[0].size[1];
+        std::vector<UMat> inputs;
+        std::vector<UMat> outputs;
+
+        inps.getUMatVector(inputs);
+        outs.getUMatVector(outputs);
+
+        int group = inputs[0].size[1] / umat_blobs[0].size[1];
 
         if (convolutionOp.empty())
         {
             OCL4DNNConvConfig config;
-            config.in_shape = shape(*inputs[0]);
+            config.in_shape = shape(inputs[0]);
             config.out_shape = shape(outputs[0]);
             config.kernel = kernel;
             config.pad = pad;
@@ -690,6 +696,112 @@ public:
             convolutionOp = Ptr<OCL4DNNConvSpatial<float> >(new OCL4DNNConvSpatial<float>(config));
         }
 
+        int k, outCn = umat_blobs[0].size[0];
+        if( weightsMat.empty() )
+        {
+            // prepare weightsMat where each row is aligned and has enough zero padding on the right to
+            // use vectorized (i.e. with intrinsics) loops without tail processing
+            Mat wm = blobs[0].reshape(1, outCn).clone();
+            if( wm.step1() % VEC_ALIGN != 0 )
+            {
+                int newcols = (int)alignSize(wm.step1(), VEC_ALIGN);
+                Mat wm_buffer = Mat(outCn, newcols, wm.type());
+                Mat wm_padding = wm_buffer.colRange(wm.cols, newcols);
+                wm_padding.setTo(Scalar::all(0.));
+                Mat wm_aligned = wm_buffer.colRange(0, wm.cols);
+                wm.copyTo(wm_aligned);
+                wm = wm_aligned;
+            }
+            weightsMat = wm;
+
+            Mat biasMat = hasBias() ? blobs[1].reshape(1, outCn) : Mat();
+            biasvec.resize(outCn+2);
+            if( biasMat.empty() )
+            {
+                for( k = 0; k < outCn; k++ )
+                    biasvec[k] = 0.f;
+            }
+            else
+            {
+                for( k = 0; k < outCn; k++ )
+                    biasvec[k] = biasMat.at<float>(k);
+            }
+
+            if( !bnorm.empty() || !scaleLayer.empty() )
+            {
+                Mat scale, shift, scale2, shift2;
+                const float *scaleptr = 0, *shiftptr = 0;
+                const float *scaleptr2 = 0, *shiftptr2 = 0;
+
+                if( !bnorm.empty() )
+                {
+                    bnorm->getScaleShift(scale, shift);
+                    CV_Assert( scale.isContinuous() && shift.isContinuous() &&
+                               scale.type() == CV_32F && shift.type() == CV_32F &&
+                               scale.total() == (size_t)outCn &&
+                               shift.total() == (size_t)outCn );
+                    scaleptr = scale.ptr<float>();
+                    shiftptr = shift.ptr<float>();
+                }
+                if( !scaleLayer.empty() )
+                {
+                    scale2 = scaleLayer->blobs[0];
+                    CV_Assert( scale2.isContinuous() && scale2.type() == CV_32F &&
+                               scale2.total() == (size_t)outCn );
+                    scaleptr2 = scale2.ptr<float>();
+                    if( scaleLayer->hasBias )
+                    {
+                        shift2 = scaleLayer->blobs[1];
+                        CV_Assert( shift2.isContinuous() && shift2.type() == CV_32F &&
+                                   shift2.total() == (size_t)outCn );
+                        shiftptr2 = shift2.ptr<float>();
+                    }
+                }
+
+                if (shiftptr || shiftptr2)
+                    fusedBias = true;
+
+                for( int i = 0; i < outCn; i++ )
+                {
+                    float s1 = scaleptr ? scaleptr[i] : 1.f;
+                    float delta1 = shiftptr ? shiftptr[i] : 0.f;
+                    float s2 = scaleptr2 ? scaleptr2[i] : 1.f;
+                    float delta2 = shiftptr2 ? shiftptr2[i] : 0.f;
+                    float* w_i = weightsMat.ptr<float>(i);
+                    int j, wcols = weightsMat.cols;
+
+                    for( j = 0; j < wcols; j++ )
+                        w_i[j] *= (s1*s2);
+
+                    biasvec[i] = biasvec[i]*(s1*s2) + (delta1*s2 + delta2);
+                }
+            }
+            biasvec[outCn] = biasvec[outCn+1] = biasvec[outCn-1];
+        }
+
+        reluslope.clear();
+        if( activ )
+        {
+            Ptr<ReLULayer> activ_relu = activ.dynamicCast<ReLULayer>();
+            if( !activ_relu.empty() )
+            {
+                reluslope.assign(outCn+2, activ_relu->negativeSlope);
+                activType = OCL4DNN_CONV_FUSED_ACTIV_RELU;
+            }
+
+            Ptr<ChannelsPReLULayer> activ_chprelu = activ.dynamicCast<ChannelsPReLULayer>();
+            if( !activ_chprelu.empty() )
+            {
+                const Mat& m = activ_chprelu->blobs[0];
+                CV_Assert(m.isContinuous() && m.type() == CV_32F && (int)m.total() == outCn);
+                const float* mdata = m.ptr<float>();
+                reluslope.resize(outCn+2);
+                std::copy(mdata, mdata + outCn, reluslope.begin());
+                reluslope[outCn] = reluslope[outCn+1] = reluslope[outCn-1];
+                activType = OCL4DNN_CONV_FUSED_ACTIV_PRELU;
+            }
+        }
+
         if ( newWeightAndBias )
         {
             weightsMat.copyTo(umat_blobs[0]);
@@ -723,9 +835,8 @@ public:
             newActiv = false;
         }
 
-        UMat inpMat, outMat;
-        inpMat = inputs[0]->getUMat(ACCESS_READ);
-        outMat = outputs[0].getUMat(ACCESS_WRITE);
+        UMat& inpMat = inputs[0];
+        UMat& outMat = outputs[0];
         int batch_size = inpMat.size[0];
 
         return convolutionOp->Forward(inpMat,
@@ -736,6 +847,18 @@ public:
     }
 #endif
 
+    void forward(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr, OutputArrayOfArrays internals_arr)
+    {
+        CV_TRACE_FUNCTION();
+        CV_TRACE_ARG_VALUE(name, "name", name.c_str());
+
+        CV_OCL_RUN((preferableTarget == DNN_TARGET_OPENCL) &&
+                   OCL_PERFORMANCE_CHECK(ocl::Device::getDefault().isIntel()),
+                   forward_ocl(inputs_arr, outputs_arr, internals_arr))
+
+        Layer::forward_fallback(inputs_arr, outputs_arr, internals_arr);
+    }
+
     void forward(std::vector<Mat*> &inputs, std::vector<Mat> &outputs, std::vector<Mat> &internals)
     {
         CV_TRACE_FUNCTION();
@@ -811,11 +934,6 @@ public:
                     }
                 }
 
-#ifdef HAVE_OPENCL
-                if (shiftptr || shiftptr2)
-                    fusedBias = true;
-#endif
-
                 for( int i = 0; i < outCn; i++ )
                 {
                     float s1 = scaleptr ? scaleptr[i] : 1.f;
@@ -841,9 +959,6 @@ public:
             if( !activ_relu.empty() )
             {
                 reluslope.assign(outCn+2, activ_relu->negativeSlope);
-#ifdef HAVE_OPENCL
-                activType = OCL4DNN_CONV_FUSED_ACTIV_RELU;
-#endif
             }
 
             Ptr<ChannelsPReLULayer> activ_chprelu = activ.dynamicCast<ChannelsPReLULayer>();
@@ -855,16 +970,9 @@ public:
                 reluslope.resize(outCn+2);
                 std::copy(mdata, mdata + outCn, reluslope.begin());
                 reluslope[outCn] = reluslope[outCn+1] = reluslope[outCn-1];
-#ifdef HAVE_OPENCL
-                activType = OCL4DNN_CONV_FUSED_ACTIV_PRELU;
-#endif
             }
         }
 
-        CV_OCL_RUN((preferableTarget == DNN_TARGET_OPENCL) &&
-                   OCL_PERFORMANCE_CHECK(ocl::Device::getDefault().isIntel()),
-                   forward_ocl(inputs, outputs, internals))
-
         int nstripes = std::max(getNumThreads(), 1);
 
         ParallelConv::run(*inputs[0], outputs[0], weightsMat, biasvec, reluslope,
@@ -1173,6 +1281,14 @@ public:
         }
     };
 
+    void forward(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr, OutputArrayOfArrays internals_arr)
+    {
+        CV_TRACE_FUNCTION();
+        CV_TRACE_ARG_VALUE(name, "name", name.c_str());
+
+        Layer::forward_fallback(inputs_arr, outputs_arr, internals_arr);
+    }
+
     void forward(std::vector<Mat *> &inputs, std::vector<Mat> &outputs, std::vector<Mat> &internals)
     {
         CV_TRACE_FUNCTION();

modules/dnn/src/layers/crop_layer.cpp

@@ -133,6 +133,14 @@ public:
         }
     }
 
+    void forward(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr, OutputArrayOfArrays internals_arr)
+    {
+        CV_TRACE_FUNCTION();
+        CV_TRACE_ARG_VALUE(name, "name", name.c_str());
+
+        Layer::forward_fallback(inputs_arr, outputs_arr, internals_arr);
+    }
+
     void forward(std::vector<Mat *> &inputs, std::vector<Mat> &outputs, std::vector<Mat> &internals)
     {
         CV_TRACE_FUNCTION();

modules/dnn/src/layers/detection_output_layer.cpp

@@ -194,6 +194,95 @@ public:
         return false;
     }
 
+#ifdef HAVE_OPENCL
+    bool forward_ocl(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr, OutputArrayOfArrays internals_arr)
+    {
+        std::vector<Mat> inpvec;
+        std::vector<Mat> outputs;
+
+        inputs_arr.getMatVector(inpvec);
+        outputs_arr.getMatVector(outputs);
+
+        std::vector<Mat*> inputs(inpvec.size());
+        for (size_t i = 0; i < inpvec.size(); i++)
+            inputs[i] = &inpvec[i];
+
+        std::vector<LabelBBox> allDecodedBBoxes;
+        std::vector<std::vector<std::vector<float> > > allConfidenceScores;
+
+        int num = inputs[0]->size[0];
+
+        // extract predictions from input layers
+        {
+            int numPriors = inputs[2]->size[2] / 4;
+
+            const float* locationData = inputs[0]->ptr<float>();
+            const float* confidenceData = inputs[1]->ptr<float>();
+            const float* priorData = inputs[2]->ptr<float>();
+
+            // Retrieve all location predictions
+            std::vector<LabelBBox> allLocationPredictions;
+            GetLocPredictions(locationData, num, numPriors, _numLocClasses,
+                              _shareLocation, _locPredTransposed, allLocationPredictions);
+
+            // Retrieve all confidences
+            GetConfidenceScores(confidenceData, num, numPriors, _numClasses, allConfidenceScores);
+
+            // Retrieve all prior bboxes
+            std::vector<caffe::NormalizedBBox> priorBBoxes;
+            std::vector<std::vector<float> > priorVariances;
+            GetPriorBBoxes(priorData, numPriors, priorBBoxes, priorVariances);
+
+            // Decode all loc predictions to bboxes
+            DecodeBBoxesAll(allLocationPredictions, priorBBoxes, priorVariances, num,
+                            _shareLocation, _numLocClasses, _backgroundLabelId,
+                            _codeType, _varianceEncodedInTarget, false, allDecodedBBoxes);
+        }
+
+        size_t numKept = 0;
+        std::vector<std::map<int, std::vector<int> > > allIndices;
+        for (int i = 0; i < num; ++i)
+        {
+            numKept += processDetections_(allDecodedBBoxes[i], allConfidenceScores[i], allIndices);
+        }
+
+        if (numKept == 0)
+        {
+            // Set confidences to zeros.
+            Range ranges[] = {Range::all(), Range::all(), Range::all(), Range(2, 3)};
+            outputs[0](ranges).setTo(0);
+            return true;
+        }
+        int outputShape[] = {1, 1, (int)numKept, 7};
+        Mat mat(4, outputShape, CV_32F);
+        float* outputsData = mat.ptr<float>();
+
+        size_t count = 0;
+        for (int i = 0; i < num; ++i)
+        {
+            count += outputDetections_(i, &outputsData[count * 7],
+                                       allDecodedBBoxes[i], allConfidenceScores[i],
+                                       allIndices[i]);
+        }
+        UMat& output = outputs_arr.getUMatRef(0);
+        output = mat.getUMat(ACCESS_READ);
+        CV_Assert(count == numKept);
+        return true;
+    }
+#endif
+
+    void forward(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr, OutputArrayOfArrays internals_arr)
+    {
+        CV_TRACE_FUNCTION();
+        CV_TRACE_ARG_VALUE(name, "name", name.c_str());
+
+        CV_OCL_RUN((preferableTarget == DNN_TARGET_OPENCL) &&
+                   OCL_PERFORMANCE_CHECK(ocl::Device::getDefault().isIntel()),
+                   forward_ocl(inputs_arr, outputs_arr, internals_arr))
+
+        Layer::forward_fallback(inputs_arr, outputs_arr, internals_arr);
+    }
+
     void forward(std::vector<Mat*> &inputs, std::vector<Mat> &outputs, std::vector<Mat> &internals)
     {
         CV_TRACE_FUNCTION();

modules/dnn/src/layers/elementwise_layers.cpp

@@ -156,13 +156,20 @@ public:
         return true;
     }
 
-    void forward(std::vector<Mat*> &inputs, std::vector<Mat> &outputs, std::vector<Mat> &internals)
+    void forward(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr, OutputArrayOfArrays internals_arr)
     {
         CV_TRACE_FUNCTION();
 
         CV_OCL_RUN((this->preferableTarget == DNN_TARGET_OPENCL) &&
                    OCL_PERFORMANCE_CHECK(ocl::Device::getDefault().isIntel()),
-                   func.applyOCL(inputs, outputs, internals))
+                   func.applyOCL(inputs_arr, outputs_arr, internals_arr))
+
+        Layer::forward_fallback(inputs_arr, outputs_arr, internals_arr);
+    }
+
+    void forward(std::vector<Mat*> &inputs, std::vector<Mat> &outputs, std::vector<Mat> &internals)
+    {
+        CV_TRACE_FUNCTION();
 
         for (size_t i = 0; i < inputs.size(); i++)
         {
@@ -258,25 +265,29 @@ struct ReLUFunctor
         return true;
     }
 
-    bool applyOCL(std::vector<Mat*> &inputs, std::vector<Mat> &outputs, std::vector<Mat> &internals)
+    bool applyOCL(InputArrayOfArrays inps, OutputArrayOfArrays outs, OutputArrayOfArrays internals)
     {
         size_t wgSize = ocl::Device::getDefault().maxWorkGroupSize();
+        std::vector<UMat> inputs;
+        std::vector<UMat> outputs;
+
+        inps.getUMatVector(inputs);
+        outs.getUMatVector(outputs);
 
         for (size_t i = 0; i < inputs.size(); i++)
         {
-            UMat src, dst;
-            inputs[i]->copyTo(src);
-            dst = outputs[i].getUMat(ACCESS_WRITE);
+            UMat& src = inputs[i];
+            UMat& dst = outputs[i];
             CV_Assert(src.isContinuous() && dst.isContinuous() && !src.offset && !dst.offset);
 
-            ocl::Kernel ker;
-            CV_Assert(initKernel(ker, src));
-            ker.set(0, (int)src.total());
-            ker.set(1, ocl::KernelArg::PtrReadOnly(src));
-            ker.set(2, ocl::KernelArg::PtrWriteOnly(dst));
+            ocl::Kernel kernel;
+            CV_Assert(initKernel(kernel, src));
+            kernel.set(0, (int)src.total());
+            kernel.set(1, ocl::KernelArg::PtrReadOnly(src));
+            kernel.set(2, ocl::KernelArg::PtrWriteOnly(dst));
 
             size_t gSize = src.total();
-            CV_Assert(ker.run(1, &gSize, &wgSize, false));
+            CV_Assert(kernel.run(1, &gSize, &wgSize, false));
         }
 
         return true;
@@ -347,7 +358,7 @@ struct ReLU6Functor
     }
 
 #ifdef HAVE_OPENCL
-    bool applyOCL(std::vector<Mat*> &inputs, std::vector<Mat> &outputs, std::vector<Mat> &internals)
+    bool applyOCL(InputArrayOfArrays inps, OutputArrayOfArrays outs, OutputArrayOfArrays internals)
     {
         // TODO: implement OCL version
         return false;
@@ -382,7 +393,7 @@ struct TanHFunctor
     }
 
 #ifdef HAVE_OPENCL
-    bool applyOCL(std::vector<Mat*> &inputs, std::vector<Mat> &outputs, std::vector<Mat> &internals)
+    bool applyOCL(InputArrayOfArrays inps, OutputArrayOfArrays outs, OutputArrayOfArrays internals)
     {
         // TODO: implement OCL version
         return false;
@@ -417,7 +428,7 @@ struct SigmoidFunctor
     }
 
 #ifdef HAVE_OPENCL
-    bool applyOCL(std::vector<Mat*> &inputs, std::vector<Mat> &outputs, std::vector<Mat> &internals)
+    bool applyOCL(InputArrayOfArrays inps, OutputArrayOfArrays outs, OutputArrayOfArrays internals)
     {
         // TODO: implement OCL version
         return false;
@@ -454,7 +465,7 @@ struct ELUFunctor
     }
 
 #ifdef HAVE_OPENCL
-    bool applyOCL(std::vector<Mat*> &inputs, std::vector<Mat> &outputs, std::vector<Mat> &internals)
+    bool applyOCL(InputArrayOfArrays inps, OutputArrayOfArrays outs, OutputArrayOfArrays internals)
     {
         // TODO: implement OCL version
         return false;
@@ -489,7 +500,7 @@ struct AbsValFunctor
     }
 
 #ifdef HAVE_OPENCL
-    bool applyOCL(std::vector<Mat*> &inputs, std::vector<Mat> &outputs, std::vector<Mat> &internals)
+    bool applyOCL(InputArrayOfArrays inps, OutputArrayOfArrays outs, OutputArrayOfArrays internals)
     {
         // TODO: implement OCL version
         return false;
@@ -524,7 +535,7 @@ struct BNLLFunctor
     }
 
 #ifdef HAVE_OPENCL
-    bool applyOCL(std::vector<Mat*> &inputs, std::vector<Mat> &outputs, std::vector<Mat> &internals)
+    bool applyOCL(InputArrayOfArrays inps, OutputArrayOfArrays outs, OutputArrayOfArrays internals)
     {
         // TODO: implement OCL version
         return false;
@@ -581,7 +592,7 @@ struct PowerFunctor
     }
 
 #ifdef HAVE_OPENCL
-    bool applyOCL(std::vector<Mat*> &inputs, std::vector<Mat> &outputs, std::vector<Mat> &internals)
+    bool applyOCL(InputArrayOfArrays inps, OutputArrayOfArrays outs, OutputArrayOfArrays internals)
     {
         // TODO: implement OCL version
         return false;
@@ -656,7 +667,7 @@ struct ChannelsPReLUFunctor
     }
 
 #ifdef HAVE_OPENCL
-    bool applyOCL(std::vector<Mat*> &inputs, std::vector<Mat> &outputs, std::vector<Mat> &internals)
+    bool applyOCL(InputArrayOfArrays inps, OutputArrayOfArrays outs, OutputArrayOfArrays internals)
     {
         // TODO: implement OCL version
         return false;

modules/dnn/src/layers/eltwise_layer.cpp

@@ -254,6 +254,14 @@ public:
         }
     };
 
+    void forward(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr, OutputArrayOfArrays internals_arr)
+    {
+        CV_TRACE_FUNCTION();
+        CV_TRACE_ARG_VALUE(name, "name", name.c_str());
+
+        Layer::forward_fallback(inputs_arr, outputs_arr, internals_arr);
+    }
+
     void forward(std::vector<Mat *> &inputs, std::vector<Mat> &outputs, std::vector<Mat> &internals)
     {
         CV_TRACE_FUNCTION();

modules/dnn/src/layers/flatten_layer.cpp

@@ -104,6 +104,43 @@ public:
         return true;
     }
 
+#ifdef HAVE_OPENCL
+    bool forward_ocl(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr, OutputArrayOfArrays internals_arr)
+    {
+        std::vector<UMat> inpvec;
+        std::vector<UMat> outputs;
+
+        inputs_arr.getUMatVector(inpvec);
+        outputs_arr.getUMatVector(outputs);
+
+        std::vector<UMat*> inputs(inpvec.size());
+        for (int i = 0; i < inpvec.size(); i++)
+            inputs[i] = &inpvec[i];
+
+        for (size_t i = 0; i < inputs.size(); i++)
+        {
+            MatShape outShape = shape(outputs[i]);
+            UMat& output = outputs_arr.getUMatRef(i);
+            output = inputs[i]->reshape(1, (int)outShape.size(), &outShape[0]);
+        }
+
+        return true;
+    }
+#endif
+
+    void forward(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr, OutputArrayOfArrays internals_arr)
+    {
+        CV_TRACE_FUNCTION();
+        CV_TRACE_ARG_VALUE(name, "name", name.c_str());
+
+        CV_OCL_RUN((preferableTarget == DNN_TARGET_OPENCL) &&
+                   outputs_arr.isUMatVector() &&
+                   OCL_PERFORMANCE_CHECK(ocl::Device::getDefault().isIntel()),
+                   forward_ocl(inputs_arr, outputs_arr, internals_arr))
+
+        Layer::forward_fallback(inputs_arr, outputs_arr, internals_arr);
+    }
+
     void forward(std::vector<Mat*> &inputs, std::vector<Mat> &outputs, std::vector<Mat> &internals)
     {
         CV_TRACE_FUNCTION();

modules/dnn/src/layers/fully_connected_layer.cpp

@@ -258,12 +258,18 @@ public:
     };
 
 #ifdef HAVE_OPENCL
-    bool forward_ocl(std::vector<Mat*> &input, std::vector<Mat> &output)
+    bool forward_ocl(InputArrayOfArrays inps, OutputArrayOfArrays outs, InputArrayOfArrays internals)
     {
-        int axisCan = clamp(axis, input[0]->dims);
-        int numOutput = blobs[0].size[0];
-        int innerSize = blobs[0].size[1];
-        int outerSize = input[0]->total(0, axisCan);
+        std::vector<UMat> inputs;
+        std::vector<UMat> outputs;
+
+        inps.getUMatVector(inputs);
+        outs.getUMatVector(outputs);
+
+        int axisCan = clamp(axis, inputs[0].dims);
+        int numOutput = umat_blobs[0].size[0];
+        int innerSize = umat_blobs[0].size[1];
+        int outerSize = total(shape(inputs[0]), 0, axisCan);
         bool ret = true;
 
         if (innerProductOp.empty())
@@ -278,11 +284,10 @@ public:
         }
 
         UMat biasOnesMat = UMat::ones(outerSize, 1, umat_blobs[0].type());
-        for (size_t i = 0; i < input.size(); i++)
+        for (size_t i = 0; i < inputs.size(); i++)
         {
-            UMat srcMat, dstMat;
-            srcMat = input[i]->reshape(1, outerSize).getUMat(ACCESS_READ);
-            dstMat = output[i].reshape(1, outerSize).getUMat(ACCESS_WRITE);
+            UMat& srcMat = inputs[i];
+            UMat& dstMat = outputs[i];
             dstMat.setTo(0.0f);
 
             if (!innerProductOp->Forward(srcMat, umat_blobs[0], (bias) ? umat_blobs[1] : UMat(), dstMat))
@@ -301,11 +306,15 @@ public:
         if (ret) return true;
 
         UMat& weights = umat_blobs[0];
-        for (size_t i = 0; i < input.size(); i++)
+        for (size_t i = 0; i < inputs.size(); i++)
         {
+            MatShape inshape, outshape;
+            inshape = shape(outerSize, innerSize);
+            outshape = shape(outerSize, numOutput);
+
             UMat srcMat, dstMat;
-            srcMat = input[i]->reshape(1, outerSize).getUMat(ACCESS_READ);
-            dstMat = output[i].reshape(1, outerSize).getUMat(ACCESS_WRITE);
+            srcMat = inputs[i].reshape(1, inshape.size(), &inshape[0]);
+            dstMat = outputs[i].reshape(1, outshape.size(), &outshape[0]);
 
             cv::gemm(srcMat, weights, 1, noArray(), 0, dstMat, GEMM_2_T);
 
@@ -320,14 +329,22 @@ public:
     }
 #endif
 
-    void forward(std::vector<Mat*> &input, std::vector<Mat> &output, std::vector<Mat> &)
+    void forward(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr, OutputArrayOfArrays internals_arr)
     {
         CV_TRACE_FUNCTION();
         CV_TRACE_ARG_VALUE(name, "name", name.c_str());
 
         CV_OCL_RUN((preferableTarget == DNN_TARGET_OPENCL) &&
                    OCL_PERFORMANCE_CHECK(ocl::Device::getDefault().isIntel()),
-                   forward_ocl(input, output))
+                   forward_ocl(inputs_arr, outputs_arr, internals_arr))
+
+        Layer::forward_fallback(inputs_arr, outputs_arr, internals_arr);
+    }
+
+    void forward(std::vector<Mat*> &input, std::vector<Mat> &output, std::vector<Mat> &)
+    {
+        CV_TRACE_FUNCTION();
+        CV_TRACE_ARG_VALUE(name, "name", name.c_str());
 
         int axisCan = clamp(axis, input[0]->dims);
         int outerSize = input[0]->total(0, axisCan);

modules/dnn/src/layers/lrn_layer.cpp

@@ -94,8 +94,14 @@ public:
     }
 
 #ifdef HAVE_OPENCL
-    bool forward_ocl(std::vector<Mat*> &inputs, std::vector<Mat> &outputs, std::vector<Mat> &internals)
+    bool forward_ocl(InputArrayOfArrays inps, OutputArrayOfArrays outs, OutputArrayOfArrays internals)
     {
+        std::vector<UMat> inputs;
+        std::vector<UMat> outputs;
+
+        inps.getUMatVector(inputs);
+        outs.getUMatVector(outputs);
+
         if (lrnOp.empty())
         {
             OCL4DNNLRNConfig config;
@@ -108,38 +114,44 @@ public:
             config.alpha = alpha;
             config.beta = beta;
             config.k = bias;
-            CHECK_EQ(4, inputs[0]->dims) << "Input must have 4 axes, "
+            CHECK_EQ(4, inputs[0].dims) << "Input must have 4 axes, "
                      << "corresponding to (num, channels, height, width)";
-            config.batch_size = inputs[0]->size[0];
-            config.channels = inputs[0]->size[1];
-            config.height = inputs[0]->size[2];
-            config.width = inputs[0]->size[3];
+            config.batch_size = inputs[0].size[0];
+            config.channels = inputs[0].size[1];
+            config.height = inputs[0].size[2];
+            config.width = inputs[0].size[3];
             config.norm_by_size = normBySize;
 
             lrnOp = Ptr<OCL4DNNLRN<float> >(new OCL4DNNLRN<float>(config));
         }
 
-        UMat inpMat, outMat;
-        inpMat = inputs[0]->getUMat(ACCESS_READ);
-        outMat = outputs[0].getUMat(ACCESS_WRITE);
-
-        if (!lrnOp->Forward(inpMat, outMat))
+        if (!lrnOp->Forward(inputs[0], outputs[0]))
             return false;
 
         return true;
     }
 #endif
 
-    void forward(std::vector<Mat*> &inputs, std::vector<Mat> &outputs, std::vector<Mat> &internals)
+    void forward(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr, OutputArrayOfArrays internals_arr)
     {
         CV_TRACE_FUNCTION();
         CV_TRACE_ARG_VALUE(name, "name", name.c_str());
 
-        CV_Assert(inputs.size() == outputs.size());
+        CV_Assert(inputs_arr.total() == outputs_arr.total());
 
         CV_OCL_RUN((preferableTarget == DNN_TARGET_OPENCL) &&
                    OCL_PERFORMANCE_CHECK(ocl::Device::getDefault().isIntel()),
-                   forward_ocl(inputs, outputs, internals))
+                   forward_ocl(inputs_arr, outputs_arr, internals_arr))
+
+        Layer::forward_fallback(inputs_arr, outputs_arr, internals_arr);
+    }
+
+    void forward(std::vector<Mat*> &inputs, std::vector<Mat> &outputs, std::vector<Mat> &internals)
+    {
+        CV_TRACE_FUNCTION();
+        CV_TRACE_ARG_VALUE(name, "name", name.c_str());
+
+        CV_Assert(inputs.size() == outputs.size());
 
         for (int i = 0; i < inputs.size(); i++)
         {

modules/dnn/src/layers/max_unpooling_layer.cpp

@@ -55,6 +55,14 @@ public:
         return false;
     }
 
+    void forward(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr, OutputArrayOfArrays internals_arr)
+    {
+        CV_TRACE_FUNCTION();
+        CV_TRACE_ARG_VALUE(name, "name", name.c_str());
+
+        Layer::forward_fallback(inputs_arr, outputs_arr, internals_arr);
+    }
+
     void forward(std::vector<Mat*> &inputs, std::vector<Mat> &outputs, std::vector<Mat> &internals)
     {
         CV_TRACE_FUNCTION();

modules/dnn/src/layers/mvn_layer.cpp

@@ -60,6 +60,14 @@ public:
         eps = params.get<double>("eps", 1e-9);
     }
 
+    void forward(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr, OutputArrayOfArrays internals_arr)
+    {
+        CV_TRACE_FUNCTION();
+        CV_TRACE_ARG_VALUE(name, "name", name.c_str());
+
+        Layer::forward_fallback(inputs_arr, outputs_arr, internals_arr);
+    }
+
     void forward(std::vector<Mat *> &inputs, std::vector<Mat> &outputs, std::vector<Mat> &internals)
     {
         CV_TRACE_FUNCTION();

modules/dnn/src/layers/normalize_bbox_layer.cpp

@@ -69,6 +69,14 @@ public:
         return true;
     }
 
+    void forward(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr, OutputArrayOfArrays internals_arr)
+    {
+        CV_TRACE_FUNCTION();
+        CV_TRACE_ARG_VALUE(name, "name", name.c_str());
+
+        Layer::forward_fallback(inputs_arr, outputs_arr, internals_arr);
+    }
+
     void forward(std::vector<Mat*> &inputs, std::vector<Mat> &outputs, std::vector<Mat> &internals)
     {
         CV_TRACE_FUNCTION();

modules/dnn/src/layers/padding_layer.cpp

@@ -91,6 +91,14 @@ public:
                backendId == DNN_BACKEND_HALIDE && haveHalide() && dstRanges.size() == 4;
     }
 
+    void forward(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr, OutputArrayOfArrays internals_arr)
+    {
+        CV_TRACE_FUNCTION();
+        CV_TRACE_ARG_VALUE(name, "name", name.c_str());
+
+        Layer::forward_fallback(inputs_arr, outputs_arr, internals_arr);
+    }
+
     void forward(std::vector<Mat*> &inputs, std::vector<Mat> &outputs, std::vector<Mat> &internals)
     {
         CV_TRACE_FUNCTION();

modules/dnn/src/layers/permute_layer.cpp

@@ -247,6 +247,14 @@ public:
         }
     };
 
+    void forward(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr, OutputArrayOfArrays internals_arr)
+    {
+        CV_TRACE_FUNCTION();
+        CV_TRACE_ARG_VALUE(name, "name", name.c_str());
+
+        Layer::forward_fallback(inputs_arr, outputs_arr, internals_arr);
+    }
+
     void forward(std::vector<Mat*> &inputs, std::vector<Mat> &outputs, std::vector<Mat> &internals)
     {
         CV_TRACE_FUNCTION();

modules/dnn/src/layers/pooling_layer.cpp

@@ -113,18 +113,24 @@ public:
     }
 
 #ifdef HAVE_OPENCL
-    bool forward_ocl(std::vector<Mat*> &inputs, std::vector<Mat> &outputs, std::vector<Mat> &internals)
+    bool forward_ocl(InputArrayOfArrays inps, OutputArrayOfArrays outs, InputArrayOfArrays internals)
     {
+        std::vector<UMat> inputs;
+        std::vector<UMat> outputs;
+
+        inps.getUMatVector(inputs);
+        outs.getUMatVector(outputs);
+
         if (poolOp.empty())
         {
             OCL4DNNPoolConfig config;
 
-            config.in_shape = shape(*inputs[0]);
+            config.in_shape = shape(inputs[0]);
             config.out_shape = shape(outputs[0]);
             config.kernel = kernel;
             config.pad = pad;
             config.stride = stride;
-            config.channels = inputs[0]->size[1];
+            config.channels = inputs[0].size[1];
             config.pool_method = type == MAX ? LIBDNN_POOLING_METHOD_MAX :
                                 (type == AVE ? LIBDNN_POOLING_METHOD_AVE :
                                                LIBDNN_POOLING_METHOD_STO);
@@ -133,18 +139,10 @@ public:
 
         for (size_t ii = 0; ii < inputs.size(); ii++)
         {
-            UMat inpMat, outMat, maskMat;
-
-            inpMat = inputs[ii]->getUMat(ACCESS_READ);
-
-            if (type == MAX)
-            {
-                outMat = outputs[2 * ii].getUMat(ACCESS_WRITE);
-                maskMat = outputs[2 * ii + 1].getUMat(ACCESS_WRITE);
-            } else {
-                outMat = outputs[ii].getUMat(ACCESS_WRITE);
-                maskMat = UMat();
-            }
+            UMat& inpMat = inputs[ii];
+            int out_index = (type == MAX) ? 2 : 1;
+            UMat& outMat = outputs[out_index * ii];
+            UMat maskMat = (type == MAX) ? outputs[2 * ii + 1] : UMat();
 
             CV_Assert(inpMat.offset == 0 && outMat.offset == 0);
 
@@ -156,14 +154,22 @@ public:
     }
 #endif
 
-    void forward(std::vector<Mat*> &inputs, std::vector<Mat> &outputs, std::vector<Mat> &internals)
+    void forward(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr, OutputArrayOfArrays internals_arr)
     {
         CV_TRACE_FUNCTION();
         CV_TRACE_ARG_VALUE(name, "name", name.c_str());
 
         CV_OCL_RUN((preferableTarget == DNN_TARGET_OPENCL) &&
                    OCL_PERFORMANCE_CHECK(ocl::Device::getDefault().isIntel()),
-                   forward_ocl(inputs, outputs, internals))
+                   forward_ocl(inputs_arr, outputs_arr, internals_arr))
+
+        Layer::forward_fallback(inputs_arr, outputs_arr, internals_arr);
+    }
+
+    void forward(std::vector<Mat*> &inputs, std::vector<Mat> &outputs, std::vector<Mat> &internals)
+    {
+        CV_TRACE_FUNCTION();
+        CV_TRACE_ARG_VALUE(name, "name", name.c_str());
 
         for (size_t ii = 0; ii < inputs.size(); ii++)
         {

modules/dnn/src/layers/prior_box_layer.cpp

@@ -249,6 +249,14 @@ public:
         return false;
     }
 
+    void forward(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr, OutputArrayOfArrays internals_arr)
+    {
+        CV_TRACE_FUNCTION();
+        CV_TRACE_ARG_VALUE(name, "name", name.c_str());
+
+        Layer::forward_fallback(inputs_arr, outputs_arr, internals_arr);
+    }
+
     void forward(std::vector<Mat*> &inputs, std::vector<Mat> &outputs, std::vector<Mat> &internals)
     {
         CV_TRACE_FUNCTION();

modules/dnn/src/layers/recurrent_layers.cpp

@@ -252,6 +252,14 @@ public:
         allocated = true;
     }
 
+    void forward(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr, OutputArrayOfArrays internals_arr)
+    {
+        CV_TRACE_FUNCTION();
+        CV_TRACE_ARG_VALUE(name, "name", name.c_str());
+
+        Layer::forward_fallback(inputs_arr, outputs_arr, internals_arr);
+    }
+
     void forward(std::vector<Mat*> &input, std::vector<Mat> &output, std::vector<Mat> &internals)
     {
         CV_TRACE_FUNCTION();
@@ -465,6 +473,14 @@ public:
         }
     }
 
+    void forward(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr, OutputArrayOfArrays internals_arr)
+    {
+        CV_TRACE_FUNCTION();
+        CV_TRACE_ARG_VALUE(name, "name", name.c_str());
+
+        Layer::forward_fallback(inputs_arr, outputs_arr, internals_arr);
+    }
+
     void forward(std::vector<Mat*> &input, std::vector<Mat> &output, std::vector<Mat> &internals)
     {
         CV_TRACE_FUNCTION();

modules/dnn/src/layers/region_layer.cpp

@@ -114,6 +114,14 @@ public:
         }
     }
 
+    void forward(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr, OutputArrayOfArrays internals_arr)
+    {
+        CV_TRACE_FUNCTION();
+        CV_TRACE_ARG_VALUE(name, "name", name.c_str());
+
+        Layer::forward_fallback(inputs_arr, outputs_arr, internals_arr);
+    }
+
     void forward(std::vector<Mat*> &inputs, std::vector<Mat> &outputs, std::vector<Mat> &internals)
     {
         CV_TRACE_FUNCTION();

modules/dnn/src/layers/reorg_layer.cpp

@@ -85,6 +85,15 @@ public:
     {
         return backendId == DNN_BACKEND_DEFAULT;
     }
+
+    void forward(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr, OutputArrayOfArrays internals_arr)
+    {
+        CV_TRACE_FUNCTION();
+        CV_TRACE_ARG_VALUE(name, "name", name.c_str());
+
+        Layer::forward_fallback(inputs_arr, outputs_arr, internals_arr);
+    }
+
     void forward(std::vector<Mat*> &inputs, std::vector<Mat> &outputs, std::vector<Mat> &internals)
     {
         CV_TRACE_FUNCTION();

modules/dnn/src/layers/reshape_layer.cpp

@@ -182,6 +182,14 @@ public:
         return true;
     }
 
+    void forward(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr, OutputArrayOfArrays internals_arr)
+    {
+        CV_TRACE_FUNCTION();
+        CV_TRACE_ARG_VALUE(name, "name", name.c_str());
+
+        Layer::forward_fallback(inputs_arr, outputs_arr, internals_arr);
+    }
+
     void forward(std::vector<Mat*> &inputs, std::vector<Mat> &outputs, std::vector<Mat> &internals)
     {
         CV_TRACE_FUNCTION();

modules/dnn/src/layers/resize_nearest_neighbor_layer.cpp

@@ -37,6 +37,14 @@ public:
         return (outputs[0][2] == inputs[0][2]) && (outputs[0][3] == inputs[0][3]);
     }
 
+    void forward(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr, OutputArrayOfArrays internals_arr)
+    {
+        CV_TRACE_FUNCTION();
+        CV_TRACE_ARG_VALUE(name, "name", name.c_str());
+
+        Layer::forward_fallback(inputs_arr, outputs_arr, internals_arr);
+    }
+
     void forward(std::vector<Mat*> &inputs, std::vector<Mat> &outputs, std::vector<Mat> &internals)
     {
         CV_TRACE_FUNCTION();

modules/dnn/src/layers/scale_layer.cpp

@@ -44,6 +44,14 @@ public:
                backendId == DNN_BACKEND_HALIDE && haveHalide();
     }
 
+    void forward(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr, OutputArrayOfArrays internals_arr)
+    {
+        CV_TRACE_FUNCTION();
+        CV_TRACE_ARG_VALUE(name, "name", name.c_str());
+
+        Layer::forward_fallback(inputs_arr, outputs_arr, internals_arr);
+    }
+
     void forward(std::vector<Mat*> &inputs, std::vector<Mat> &outputs, std::vector<Mat> &internals)
     {
         CV_TRACE_FUNCTION();

modules/dnn/src/layers/shift_layer.cpp

@@ -36,6 +36,14 @@ public:
         return true;
     }
 
+    void forward(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr, OutputArrayOfArrays internals_arr)
+    {
+        CV_TRACE_FUNCTION();
+        CV_TRACE_ARG_VALUE(name, "name", name.c_str());
+
+        Layer::forward_fallback(inputs_arr, outputs_arr, internals_arr);
+    }
+
     virtual void forward(std::vector<Mat*> &inputs, std::vector<Mat> &outputs, std::vector<Mat> &internals)
     {
         CV_TRACE_FUNCTION();

modules/dnn/src/layers/slice_layer.cpp

@@ -171,6 +171,14 @@ public:
         }
     }
 
+    void forward(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr, OutputArrayOfArrays internals_arr)
+    {
+        CV_TRACE_FUNCTION();
+        CV_TRACE_ARG_VALUE(name, "name", name.c_str());
+
+        Layer::forward_fallback(inputs_arr, outputs_arr, internals_arr);
+    }
+
     void forward(std::vector<Mat*> &inputs, std::vector<Mat> &outputs, std::vector<Mat> &internals)
     {
         CV_TRACE_FUNCTION();

modules/dnn/src/layers/softmax_layer.cpp

@@ -91,35 +91,42 @@ public:
     }
 
 #ifdef HAVE_OPENCL
-    bool forward_ocl(std::vector<Mat*> &inputs, std::vector<Mat> &outputs, std::vector<Mat> &internals)
+    bool forward_ocl(InputArrayOfArrays inps, OutputArrayOfArrays outs, OutputArrayOfArrays itns)
     {
+        std::vector<UMat> inputs;
+        std::vector<UMat> outputs;
+        std::vector<UMat> internals;
+
+        inps.getUMatVector(inputs);
+        outs.getUMatVector(outputs);
+        itns.getUMatVector(internals);
+
         if (softmaxOp.empty())
         {
             OCL4DNNSoftmaxConfig config;
 
-            config.in_shape = shape(*inputs[0]);
+            config.in_shape = shape(inputs[0]);
             config.axis = axisRaw;
-            config.channels = inputs[0]->size[axisRaw];
+            config.channels = inputs[0].size[axisRaw];
             config.logsoftmax = logSoftMax;
 
             softmaxOp = Ptr<OCL4DNNSoftmax<float> >(new OCL4DNNSoftmax<float>(config));
         }
 
-        UMat srcMat, dstMat;
-        srcMat = inputs[0]->getUMat(ACCESS_READ);
-        dstMat = outputs[0].getUMat(ACCESS_WRITE);
+        UMat& src = inputs[0];
+        UMat& dstMat = outputs[0];
 
-        if (softmaxOp->Forward(srcMat, dstMat))
+        if (softmaxOp->Forward(src, dstMat))
             return true;
 
-        const Mat &src = *inputs[0];
-        UMat bufMat = internals[0].getUMat(ACCESS_WRITE);
-        srcMat.copyTo(dstMat);
+        UMat& bufMat = internals[0];
+        src.copyTo(dstMat);
 
         int axis = clamp(axisRaw, src.dims);
-        size_t outerSize = src.total(0, axis);
+        MatShape s = shape(src);
+        size_t outerSize = total(s, 0, axis);
         size_t channels = src.size[axis];
-        size_t innerSize = src.total(axis + 1);
+        size_t innerSize = total(s, axis + 1);
 
         String buildOpts = String("-DT=") + ocl::typeToStr(src.type());
         ocl::Kernel kmax, ksub, ksum, kdiv;
@@ -175,14 +182,22 @@ public:
     }
 #endif
 
-    void forward(std::vector<Mat*> &inputs, std::vector<Mat> &outputs, std::vector<Mat> &internals)
+    void forward(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr, OutputArrayOfArrays internals_arr)
     {
         CV_TRACE_FUNCTION();
         CV_TRACE_ARG_VALUE(name, "name", name.c_str());
 
         CV_OCL_RUN((preferableTarget == DNN_TARGET_OPENCL) &&
                    OCL_PERFORMANCE_CHECK(ocl::Device::getDefault().isIntel()),
-                   forward_ocl(inputs, outputs, internals))
+                   forward_ocl(inputs_arr, outputs_arr, internals_arr))
+
+        Layer::forward_fallback(inputs_arr, outputs_arr, internals_arr);
+    }
+
+    void forward(std::vector<Mat*> &inputs, std::vector<Mat> &outputs, std::vector<Mat> &internals)
+    {
+        CV_TRACE_FUNCTION();
+        CV_TRACE_ARG_VALUE(name, "name", name.c_str());
 
         const Mat &src = *inputs[0];
         Mat &dst = outputs[0];

modules/dnn/src/layers/split_layer.cpp

@@ -78,6 +78,14 @@ public:
         return false;
     }
 
+    void forward(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr, OutputArrayOfArrays internals_arr)
+    {
+        CV_TRACE_FUNCTION();
+        CV_TRACE_ARG_VALUE(name, "name", name.c_str());
+
+        Layer::forward_fallback(inputs_arr, outputs_arr, internals_arr);
+    }
+
     void forward(std::vector<Mat*> &inputs, std::vector<Mat> &outputs, std::vector<Mat> &internals)
     {
         CV_TRACE_FUNCTION();