dnn fp16 support

Signed-off-by: NLi Peng <peng.li@intel.com>

modules/dnn/src/dnn.cpp

@@ -499,7 +499,7 @@ public:
         }
     }
 
-    void reuseOrCreate(const MatShape& shape, const LayerPin& lp, Mat& dst, bool forceCreate)
+    void reuseOrCreate(const MatShape& shape, const LayerPin& lp, Mat& dst, bool forceCreate, bool use_half)
     {
         if (!DNN_DISABLE_MEMORY_OPTIMIZATIONS && !forceCreate)
         {
@@ -540,14 +540,14 @@ public:
         {
             // if dst already has been allocated with total(shape) elements,
             // it won't be recrreated and pointer of dst.data remains the same.
-            dst.create(shape, CV_32F);
+            dst.create(shape, use_half ? CV_16S : CV_32F);
             addHost(lp, dst);
         }
     }
 
     void allocateBlobsForLayer(LayerData &ld, const LayerShapes& layerShapes,
                                std::vector<LayerPin>& pinsForInternalBlobs,
-                               bool forceCreate = false)
+                               bool forceCreate = false, bool use_half = false)
     {
         CV_TRACE_FUNCTION();
 
@@ -618,7 +618,7 @@ public:
                         reuse(ld.inputBlobsId[0], blobPin);
                     }
                     else
-                        reuseOrCreate(shapes[index], blobPin, *blobs[index], forceCreate);
+                        reuseOrCreate(shapes[index], blobPin, *blobs[index], forceCreate, use_half);
                 }
             }
         }
@@ -656,7 +656,7 @@ static Ptr<BackendWrapper> wrapMat(int backendId, int targetId, cv::Mat& m)
     {
         if (targetId == DNN_TARGET_CPU)
             return Ptr<BackendWrapper>();
-        else if (targetId == DNN_TARGET_OPENCL)
+        else if (IS_DNN_OPENCL_TARGET(targetId))
             return OpenCLBackendWrapper::create(m);
         else
             CV_Error(Error::StsNotImplemented, "Unknown target identifier");
@@ -721,6 +721,7 @@ struct Net::Impl
     bool netWasAllocated;
     bool fusion;
     std::vector<int64> layersTimings;
+    Mat output_blob;
 
     Ptr<BackendWrapper> wrap(Mat& host)
     {
@@ -737,7 +738,7 @@ struct Net::Impl
             Ptr<BackendWrapper> baseBuffer = backendWrappers[data];
             if (preferableBackend == DNN_BACKEND_DEFAULT)
             {
-                CV_Assert(preferableTarget == DNN_TARGET_OPENCL);
+                CV_Assert(IS_DNN_OPENCL_TARGET(preferableTarget));
                 return OpenCLBackendWrapper::create(baseBuffer, host);
             }
             else if (preferableBackend == DNN_BACKEND_HALIDE)
@@ -849,7 +850,7 @@ struct Net::Impl
 
         if (!netWasAllocated || this->blobsToKeep != blobsToKeep_)
         {
-            if (preferableBackend == DNN_BACKEND_DEFAULT && preferableTarget == DNN_TARGET_OPENCL)
+            if (preferableBackend == DNN_BACKEND_DEFAULT && IS_DNN_OPENCL_TARGET(preferableTarget))
 #ifndef HAVE_OPENCL
             {
                 CV_LOG_WARNING(NULL, "DNN: OpenCL target is not available in this OpenCV build, switching to CPU.");
@@ -1034,7 +1035,7 @@ struct Net::Impl
     {
         CV_TRACE_FUNCTION();
         if (preferableBackend == DNN_BACKEND_DEFAULT)
-            CV_Assert(preferableTarget == DNN_TARGET_CPU || preferableTarget == DNN_TARGET_OPENCL);
+            CV_Assert(preferableTarget == DNN_TARGET_CPU || IS_DNN_OPENCL_TARGET(preferableTarget));
         else if (preferableBackend == DNN_BACKEND_HALIDE)
             initHalideBackend();
         else if (preferableBackend == DNN_BACKEND_INFERENCE_ENGINE)
@@ -1369,7 +1370,9 @@ struct Net::Impl
 
         std::vector<LayerPin> pinsForInternalBlobs;
         blobManager.allocateBlobsForLayer(ld, layerShapesIt->second, pinsForInternalBlobs,
-                                          preferableBackend == DNN_BACKEND_INFERENCE_ENGINE);
+                                          preferableBackend == DNN_BACKEND_INFERENCE_ENGINE,
+                                          preferableBackend == DNN_BACKEND_DEFAULT &&
+                                          preferableTarget == DNN_TARGET_OPENCL_FP16);
         ld.outputBlobsWrappers.resize(ld.outputBlobs.size());
         for (int i = 0; i < ld.outputBlobs.size(); ++i)
         {
@@ -1439,7 +1442,7 @@ struct Net::Impl
             // some other layers.
 
             // TODO: OpenCL target support more fusion styles.
-            if ( preferableBackend == DNN_BACKEND_DEFAULT && preferableTarget == DNN_TARGET_OPENCL &&
+            if ( preferableBackend == DNN_BACKEND_DEFAULT && IS_DNN_OPENCL_TARGET(preferableTarget) &&
                  (!cv::ocl::useOpenCL() || (ld.layerInstance->type != "Convolution" &&
                  ld.layerInstance->type != "MVN")) )
                 continue;
@@ -1478,8 +1481,8 @@ struct Net::Impl
                     continue;  // Go to the next layer.
 
                 // For now, OpenCL target support fusion with activation of ReLU/ChannelsPReLU/Power/Tanh
-                if ( preferableTarget != DNN_TARGET_OPENCL ||
-                        (preferableTarget == DNN_TARGET_OPENCL &&
+                if ( !IS_DNN_OPENCL_TARGET(preferableTarget) ||
+                     (IS_DNN_OPENCL_TARGET(preferableTarget) &&
                          nextData &&
                         ((nextData->type == "ReLU") ||
                          (nextData->type == "ChannelsPReLU") ||
@@ -1502,7 +1505,7 @@ struct Net::Impl
                         ld.outputBlobs = layers[lpNext.lid].outputBlobs;
                         ld.outputBlobsWrappers = layers[lpNext.lid].outputBlobsWrappers;
 
-                        if ( preferableTarget == DNN_TARGET_OPENCL )
+                        if ( IS_DNN_OPENCL_TARGET(preferableTarget) )
                         {
                             if ( !activData->consumers.empty() )
                             {
@@ -1514,7 +1517,7 @@ struct Net::Impl
                 }
 
                 // fuse convlution layer followed by eltwise + relu
-                if ( preferableTarget == DNN_TARGET_OPENCL )
+                if ( IS_DNN_OPENCL_TARGET(preferableTarget) )
                 {
                     Ptr<EltwiseLayer> nextEltwiseLayer;
                     if( nextData )
@@ -1727,6 +1730,13 @@ struct Net::Impl
         for(int i = 0; i < layers[0].outputBlobs.size(); i++)
         {
             CV_Assert(layers[0].outputBlobs[i].total());
+            if (layers[0].outputBlobs[i].depth() == CV_32F &&
+                preferableBackend == DNN_BACKEND_DEFAULT &&
+                preferableTarget == DNN_TARGET_OPENCL_FP16)
+            {
+                Mat mat = layers[0].outputBlobs[i].clone();
+                convertFp16(mat, layers[0].outputBlobs[i]);
+            }
             inputShapes.push_back(shape(layers[0].outputBlobs[i]));
         }
         LayersShapesMap layersShapes;
@@ -1772,7 +1782,7 @@ struct Net::Impl
         {
             if( !ld.skip )
             {
-                if (preferableBackend == DNN_BACKEND_DEFAULT && preferableTarget == DNN_TARGET_OPENCL)
+                if (preferableBackend == DNN_BACKEND_DEFAULT && IS_DNN_OPENCL_TARGET(preferableTarget))
                 {
                     std::vector<UMat> umat_outputBlobs = OpenCLBackendWrapper::getUMatVector(ld.outputBlobsWrappers);
                     layer->forward(OpenCLBackendWrapper::getUMatVector(ld.inputBlobsWrappers),
@@ -1937,7 +1947,14 @@ struct Net::Impl
             // Transfer data to CPU if it's require.
             ld.outputBlobsWrappers[pin.oid]->copyToHost();
         }
-        return ld.outputBlobs[pin.oid];
+
+        if (ld.outputBlobs[pin.oid].depth() == CV_16S)
+        {
+            convertFp16(ld.outputBlobs[pin.oid], output_blob);
+            return output_blob;
+        }
+        else
+            return ld.outputBlobs[pin.oid];
     }
 
     Mat getBlob(String outputName)
@@ -2080,7 +2097,7 @@ void Net::forward(OutputArrayOfArrays outputBlobs, const String& outputName)
 
     if (outputBlobs.isUMat())
     {
-        outputBlobs.assign(ld.outputBlobs[pin.oid].getUMat(ACCESS_RW));
+        outputBlobs.assign(impl->getBlob(layerName).getUMat(ACCESS_RW));
     }
     else if (outputBlobs.isMat())
     {
@@ -2096,17 +2113,33 @@ void Net::forward(OutputArrayOfArrays outputBlobs, const String& outputName)
                 ld.outputBlobsWrappers[i]->copyToHost();
             }
         }
-        std::vector<Mat> & outputvec = *(std::vector<Mat> *)outputBlobs.getObj();
-        outputvec = ld.outputBlobs;
+        if (ld.outputBlobs[0].depth() == CV_32F)
+        {
+            std::vector<Mat> & outputvec = *(std::vector<Mat> *)outputBlobs.getObj();
+            outputvec = ld.outputBlobs;
+        } else {
+            std::vector<Mat> & outputvec = *(std::vector<Mat> *)outputBlobs.getObj();
+            outputvec.resize(ld.outputBlobs.size());
+            for (int i = 0; i < outputvec.size(); i++)
+                convertFp16(ld.outputBlobs[i], outputvec[i]);
+        }
     }
     else if (outputBlobs.isUMatVector())
     {
         std::vector<UMat> & outputvec = *(std::vector<UMat> *)outputBlobs.getObj();
 
         if (impl->preferableBackend == DNN_BACKEND_DEFAULT &&
-            impl->preferableTarget == DNN_TARGET_OPENCL)
+            IS_DNN_OPENCL_TARGET(impl->preferableTarget))
         {
-            outputvec = OpenCLBackendWrapper::getUMatVector(ld.outputBlobsWrappers);
+            if (impl->preferableTarget == DNN_TARGET_OPENCL)
+                outputvec = OpenCLBackendWrapper::getUMatVector(ld.outputBlobsWrappers);
+            else if (impl->preferableTarget == DNN_TARGET_OPENCL_FP16)
+            {
+                std::vector<UMat> out_vec = OpenCLBackendWrapper::getUMatVector(ld.outputBlobsWrappers);
+                outputvec.resize(out_vec.size());
+                for (int i = 0; i < out_vec.size(); i++)
+                    convertFp16(out_vec[i], outputvec[i]);
+            }
         }
         else
         {
@@ -2194,6 +2227,16 @@ void Net::setPreferableTarget(int targetId)
     if( impl->preferableTarget != targetId )
     {
         impl->preferableTarget = targetId;
+        if (IS_DNN_OPENCL_TARGET(targetId))
+        {
+#ifndef HAVE_OPENCL
+            impl->preferableTarget = DNN_TARGET_CPU;
+#else
+            bool fp16 = ocl::Device::getDefault().isExtensionSupported("cl_khr_fp16");
+            if (!fp16 && targetId == DNN_TARGET_OPENCL_FP16)
+                impl->preferableTarget = DNN_TARGET_OPENCL;
+#endif
+        }
         impl->netWasAllocated = false;
         impl->clear();
     }
@@ -2222,7 +2265,17 @@ void Net::setInput(InputArray blob, const String& name)
     ld.outputBlobs.resize( std::max(pin.oid+1, (int)ld.requiredOutputs.size()) );
     ld.outputBlobsWrappers.resize(ld.outputBlobs.size());
     MatShape prevShape = shape(ld.outputBlobs[pin.oid]);
-    Mat blob_ = blob.getMat();
+    Mat blob_;
+    if (impl->preferableBackend == DNN_BACKEND_DEFAULT &&
+        impl->preferableTarget == DNN_TARGET_OPENCL_FP16)
+    {
+        Mat blob_mat = blob.getMat();
+        convertFp16(blob_mat, blob_);
+    }
+    else
+    {
+        blob_ = blob.getMat();
+    }
     bool oldShape = prevShape == shape(blob_);
     if (oldShape)
     {
@@ -2747,6 +2800,43 @@ void Layer::forward_fallback(InputArrayOfArrays inputs_arr, OutputArrayOfArrays
     CV_TRACE_FUNCTION();
     CV_TRACE_ARG_VALUE(name, "name", name.c_str());
 
+    if (preferableTarget == DNN_TARGET_OPENCL_FP16 && inputs_arr.depth() == CV_16S)
+    {
+        std::vector<UMat> inputs;
+        std::vector<UMat> outputs;
+        std::vector<UMat> internals;
+
+        std::vector<UMat> orig_inputs;
+        std::vector<UMat> orig_outputs;
+        std::vector<UMat> orig_internals;
+
+        inputs_arr.getUMatVector(orig_inputs);
+        outputs_arr.getUMatVector(orig_outputs);
+        internals_arr.getUMatVector(orig_internals);
+
+        inputs.resize(orig_inputs.size());
+        for (size_t i = 0; i < orig_inputs.size(); i++)
+            convertFp16(orig_inputs[i], inputs[i]);
+
+        outputs.resize(orig_outputs.size());
+        for (size_t i = 0; i < orig_outputs.size(); i++)
+            outputs[i].create(shape(orig_outputs[i]), CV_32F);
+
+        internals.resize(orig_internals.size());
+        for (size_t i = 0; i < orig_internals.size(); i++)
+            internals[i].create(shape(orig_internals[i]), CV_32F);
+
+        forward(inputs, outputs, internals);
+
+        for (size_t i = 0; i < outputs.size(); i++)
+            convertFp16(outputs[i], orig_outputs[i]);
+
+        // sync results back
+        outputs_arr.assign(orig_outputs);
+        internals_arr.assign(orig_internals);
+        return;
+    }
+
     std::vector<Mat> inpvec;
     std::vector<Mat> outputs;
     std::vector<Mat> internals;

modules/dnn/src/precomp.hpp

@@ -64,6 +64,7 @@
 
 namespace cv { namespace dnn {
 CV__DNN_EXPERIMENTAL_NS_BEGIN
+#define IS_DNN_OPENCL_TARGET(id) (id == DNN_TARGET_OPENCL || id == DNN_TARGET_OPENCL_FP16)
 Mutex& getInitializationMutex();
 void initializeLayerFactory();
 CV__DNN_EXPERIMENTAL_NS_END