From 67f9406cbe32b967ac9d5dbcb6de96866dceb81b Mon Sep 17 00:00:00 2001
From: Li Peng <peng.li@intel.com>
Date: Thu, 4 Jan 2018 19:47:35 +0800
Subject: [PATCH] add normalize_bbox layer ocl implementation

Signed-off-by: Li Peng <peng.li@intel.com>
---
 modules/dnn/src/layers/normalize_bbox_layer.cpp | 63 +++++++++++++++++++++++++
 modules/dnn/test/test_torch_importer.cpp        |  5 ++
 2 files changed, 68 insertions(+)

diff --git a/modules/dnn/src/layers/normalize_bbox_layer.cpp b/modules/dnn/src/layers/normalize_bbox_layer.cpp
index 9d2940d..4344589 100644
--- a/modules/dnn/src/layers/normalize_bbox_layer.cpp
+++ b/modules/dnn/src/layers/normalize_bbox_layer.cpp
@@ -69,11 +69,74 @@ public:
         return true;
     }
 
+#ifdef HAVE_OPENCL
+    bool forward_ocl(InputArrayOfArrays inputs_, OutputArrayOfArrays outputs_, OutputArrayOfArrays internals_)
+    {
+        std::vector<UMat> inputs;
+        std::vector<UMat> outputs;
+        std::vector<UMat> internals;
+
+        inputs_.getUMatVector(inputs);
+        outputs_.getUMatVector(outputs);
+        internals_.getUMatVector(internals);
+
+        CV_Assert(inputs.size() == 1 && outputs.size() == 1);
+        CV_Assert(inputs[0].total() == outputs[0].total());
+
+        const UMat& inp0 = inputs[0];
+        UMat& buffer = internals[0];
+        size_t num = inp0.size[0];
+        size_t channels = inp0.size[1];
+        size_t channelSize = inp0.total() / (num * channels);
+        for (size_t i = 0; i < num; ++i)
+        {
+            MatShape s = shape(channels, channelSize);
+            UMat src = inputs[i].reshape(1, s.size(), &s[0]);
+            UMat dst = outputs[i].reshape(1, s.size(), &s[0]);
+
+            UMat abs_mat;
+            absdiff(src, cv::Scalar::all(0), abs_mat);
+            pow(abs_mat, pnorm, buffer);
+
+            if (acrossSpatial)
+            {
+                // add eps to avoid overflow
+                float absSum = sum(buffer)[0] + epsilon;
+                float norm = pow(absSum, 1.0f / pnorm);
+                multiply(src, 1.0f / norm, dst);
+            }
+
+            if (!blobs.empty())
+            {
+                // scale the output
+                Mat scale = blobs[0];
+                if (scale.total() == 1)
+                {
+                    // _scale: 1 x 1
+                    multiply(dst, scale.at<float>(0, 0), dst);
+                }
+                else
+                {
+                    // _scale: _channels x 1
+                    CV_Assert(scale.total() == channels);
+                    repeat(scale, 1, dst.cols, buffer);
+                    multiply(dst, buffer, dst);
+                }
+            }
+        }
+        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);
     }
 
diff --git a/modules/dnn/test/test_torch_importer.cpp b/modules/dnn/test/test_torch_importer.cpp
index cbac6c5..8ae0ca2 100644
--- a/modules/dnn/test/test_torch_importer.cpp
+++ b/modules/dnn/test/test_torch_importer.cpp
@@ -225,6 +225,11 @@ TEST(Torch_Importer, net_normalize)
     runTorchNet("net_normalize", DNN_TARGET_CPU, "", false, true);
 }
 
+OCL_TEST(Torch_Importer, net_normalize)
+{
+    runTorchNet("net_normalize", DNN_TARGET_OPENCL, "", false, true);
+}
+
 TEST(Torch_Importer, net_padding)
 {
     runTorchNet("net_padding", DNN_TARGET_CPU, "", false, true);
-- 
2.7.4