load fp16 as fp32 and align fp16 and double in onnx_graph_simplifie

modules/dnn/src/onnx/onnx_graph_simplifier.cpp

@@ -18,6 +18,17 @@ CV__DNN_INLINE_NS_BEGIN
 
 extern bool DNN_DIAGNOSTICS_RUN;
 
+static int isLittleEndianCPU()
+{
+    int x = 7;
+    char *ptr = (char *)&x;
+
+    if(ptr[0] == 0)
+        return 0;
+    else
+        return 1;
+}
+
 // This wrapper can behave differently for fake input nodes and real graph nodes.
 class ONNXNodeWrapper : public ImportNodeWrapper
 {
@@ -767,11 +778,64 @@ Mat getMatFromTensor(const opencv_onnx::TensorProto& tensor_proto)
             Mat(sizes, CV_32FC1, val).copyTo(blob);
         }
     }
+    else if (datatype == opencv_onnx::TensorProto_DataType_FLOAT16)
+    {
+        // FIXME, for now, we only load FP16 Tensor as FP32 Mat, full support for FP16 is required in the future.
+        CV_LOG_ONCE_WARNING(NULL, "DNN: load FP16 model as FP32 model, and it takes twice the FP16 RAM requirement.");
+
+        // ONNX saves float 16 data in two format: int32 and raw_data.
+        // Link: https://github.com/onnx/onnx/issues/4460#issuecomment-1224373746
+        if (!tensor_proto.int32_data().empty())
+        {
+            const int offset = isLittleEndianCPU() ? 0 : 1;
+            const ::google::protobuf::RepeatedField<int32_t> field = tensor_proto.int32_data();
+
+            AutoBuffer<float16_t, 16> aligned_val;
+            size_t sz = tensor_proto.int32_data().size();
+            aligned_val.allocate(sz);
+            float16_t* bufPtr = aligned_val.data();
+
+            float16_t *fp16Ptr = (float16_t *)field.data();
+            for (int i = 0; i < sz; i++)
+            {
+                bufPtr[i] = fp16Ptr[i*2 + offset];
+            }
+            Mat(sizes, CV_16FC1, bufPtr).convertTo(blob, CV_32FC1);
+        }
+        else
+        {
+            char* val = const_cast<char*>(tensor_proto.raw_data().c_str());
+#if CV_STRONG_ALIGNMENT
+            // Aligned pointer is required.
+            AutoBuffer<float16_t, 16> aligned_val;
+            if (!isAligned<sizeof(float16_t)>(val))
+            {
+                size_t sz = tensor_proto.raw_data().size();
+                aligned_val.allocate(divUp(sz, sizeof(float16_t)));
+                memcpy(aligned_val.data(), val, sz);
+                val = (char*)aligned_val.data();
+            }
+#endif
+            Mat(sizes, CV_16FC1, val).convertTo(blob, CV_32FC1);
+        }
+    }
     else if (datatype == opencv_onnx::TensorProto_DataType_DOUBLE)
     {
         const ::google::protobuf::RepeatedField<double> field = tensor_proto.double_data();
         CV_Assert(!field.empty());
-        Mat(sizes, CV_64FC1, (void*)field.data()).convertTo(blob, CV_32FC1);
+        char* val = (char *)field.data();
+#if CV_STRONG_ALIGNMENT
+        // Aligned pointer is required.
+        AutoBuffer<double, 16> aligned_val;
+        if (!isAligned<sizeof(double)>(val))
+        {
+            size_t sz = tensor_proto.raw_data().size();
+            aligned_val.allocate(divUp(sz, sizeof(double)));
+            memcpy(aligned_val.data(), val, sz);
+            val = (char*)aligned_val.data();
+        }
+#endif
+        Mat(sizes, CV_64FC1, val).convertTo(blob, CV_32FC1);
     }
     else if (datatype == opencv_onnx::TensorProto_DataType_INT32)
     {

modules/dnn/test/test_onnx_importer.cpp

@@ -2098,6 +2098,11 @@ TEST_P(Test_ONNX_nets, MobileNet_v2)
     testONNXModels("mobilenetv2", pb, default_l1, default_lInf, true);
 }
 
+TEST_P(Test_ONNX_nets, MobileNet_v2_FP16)
+{
+    testONNXModels("mobilenetv2_fp16", npy, default_l1, default_lInf, true);
+}
+
 TEST_P(Test_ONNX_nets, LResNet100E_IR)
 {
     applyTestTag(