In a nutshell, I need help with the right use of unique_ptr and not with the library ArmNN. So, the next paragraph is just for contextualization.
I am adapting my current application to use the library ArmNN. More specifically, I am doing that through the use of the interface ICaffeParser.
At line 22 of this interface, we have this using definition to define a unique_ptr to the interface, that I believe is the "cause" of my problems.
using ICaffeParserPtr = std::unique_ptr<ICaffeParser, void(*)(ICaffeParser* parser)>;
I am quite sure my problem is the incorrect use of unique_ptr in my context, once I could make some successful tests with a more simple application.
My current code contains a class, let's call it MyClass:
namespace MYNAMESPACE {
class MyClass {
public:
MyClass() {
}
// a lot of functions
// a lot of attributes
private:
// a lot of functions
// a lot of attributes
}
}
In order to make use of the ArmNN library, I have created a new private attribute for MyClass:
armnnCaffeParser::ICaffeParserPtr myParser;
and instantiated myParser at MyClass() constructor:
MyClass::MyClass() {
myParser = armnnCaffeParser::ICaffeParser::Create();
}
Remembering ICaffeParserPtr is a unique_ptr (I think), now I have the following compiling error:
/my_path/src/detector.cpp: In constructor ‘MYNAMESPACE::MyClass::MyClass()’:
/my_path/src/detector.cpp:13:20: error: no matching function for call to ‘std::unique_ptr<armnnCaffeParser::ICaffeParser, void (*)(armnnCaffeParser::ICaffeParser*)>::unique_ptr()’
MyClass::MyClass() {
^
In file included from /usr/aarch64-linux-gnu/include/c++/7/bits/locale_conv.h:41:0,
from /usr/aarch64-linux-gnu/include/c++/7/locale:43,
from /usr/aarch64-linux-gnu/include/c++/7/iomanip:43,
from /usr/include/opencv2/flann/lsh_table.h:40,
from /usr/include/opencv2/flann/lsh_index.h:49,
from /usr/include/opencv2/flann/all_indices.h:42,
from /usr/include/opencv2/flann/flann_base.hpp:43,
from /usr/include/opencv2/flann.hpp:48,
from /usr/include/opencv2/opencv.hpp:62,
from /my_path/src/detector.hpp:11,
from /my_path/src/detector.cpp:1:
/usr/aarch64-linux-gnu/include/c++/7/bits/unique_ptr.h:255:2: note: candidate: template<class _Up, class> std::unique_ptr<_Tp, _Dp>::unique_ptr(std::auto_ptr<_Up>&&)
unique_ptr(auto_ptr<_Up>&& __u) noexcept;
/usr/aarch64-linux-gnu/include/c++/7/bits/unique_ptr.h:255:2: note: template argument deduction/substitution failed:
/my_path/src/detector.cpp:13:20: note: candidate expects 1 argument, 0 provided
MyClass::MyClass() {
^
The error happens because myParser is actually being default-initialized and then assigned on the constructor body of MyClass::MyClass().
Since a function pointer is passed as a custom deleter to std::unique_ptr to form the ICaffeParserPtr type, the default constructor for this particular instance of std::unique_ptr is disabled as per [unique.ptr.single.ctor].
In other words, ICaffeParserPtr, for safety reasons, cannot be default-initialized — which specific function to otherwise assign as its deleter on initialization?
To address this, you should always initialize class members at the member initializer list. In this case, initialize myParser as such:
MyClass::MyClass():
myParser(armnnCaffeParser::ICaffeParser::Create()) {}
This avoids calling the unavailable default constructor for std::unique_ptr, and is generally a better practice than assigning to class members in the constructor body.
Related
Both Java and Javascript allow for a different way of executing static code. Java allows you to have static code in the body of a class while JS allows you to execute static code outside class definitions. Examples:
Java:
public class MyClass {
private static Map<String,String> someMap = new HashMap<String,String();
static {
someMap.put("key1","value");
someMap.put("key2","value");
SomeOtherClass.someOtherStaticMethod();
System.out.println(someMap);
}
}
JS (basically any JS code outside a class):
var myint = 5;
callSomeMethod();
$(document).ready(function () {
$("#hiddenelement").hide();
});
However, it seems like Dart supports either of both ways. Declaring global variables and methods is supported, but calling methods and executing code like in JS is not. This can only be done in a main() method. Also, static code inside a class is not allowed either.
I know Dart has other ways to statically fill a Map like my first example, but there is another case that I can think of for which this is required.
Let's consider the following CarRegistry implementation that allows you to map strings of the car model to an instance of the corresponding class. F.e. when you get the car models from JSON data:
class CarRegistry {
static Map<String, Function> _factoryMethods = new HashMap<String, Function>();
static void registerFactory(String key, Car factory()) {
_factoryMethods[key] = factory;
}
static Car createInstance(String key) {
Function factory = _factoryMethods[key];
if(factory != null) {
return factory();
}
throw new Exception("Key not found: $key");
}
}
class TeslaModelS extends Car {
}
class TeslaModelX extends Car {
}
In order to be able to call CarRegistry.createInstance("teslamodelx");, the class must first be registered. In Java this could be done by adding the following line to each Car class: static { CarRegistry.registerFactory("teslamodelx" , () => new TeslaModelX()); }. What you don't want is to hard-code all cars into the registry, because it will lose it's function as a registry, and it increases coupling. You want to be able to add a new car by only adding one new file. In JS you could call the CarRegistry.registerFactory("teslamodelx" , () => new TeslaModelX()); line outside the class construct.
How could a similar thing be done in Dart?
Even if you would allow to edit multiple files to add a new car, it would not be possible if you are writing a library without a main() method. The only option then is to fill the map on the first call of the Registry.createInstance() method, but it's no longer a registry then, just a class containing a hard-coded list of cars.
EDIT: A small addition to the last statement I made here: filling this kind of registry in the createInstance() method is only an option if the registry resided in my own library. If, f.e. I want to register my own classes to a registry provided by a different library that I imported, that's no longer an option.
Why all the fuss about static?
You can create a getter that checks if the initialization was already done (_factoryMethods != null) if not do it and return the map.
As far a I understand it, this is all about at what time this code should be executed.
The approach I showed above is lazy initialization.
I think this is usually the preferred way I guess.
If you want to do initialization when the library is loaded
I don't know another way as calling an init() method of the library from main() and add initialization code this libraries init() method.
Here is a discussion about this topic executing code at library initialization time
I encountered the same issue when trying to drive a similarly themed library.
My initial attempt explored using dart:mirrors to iterate over classes in a library and determine if they were tagged by an annotation like this (using your own code as part of the example):
#car('telsamodelx')
class TelsaModelX extends Car {
}
If so, they got automatically populated into the registry. Performance wasn't great, though, and I wasn't sure if how it was going to scale.
I ended up taking a more cumbersome approach:
// Inside of CarRegistry.dart
class CarRegister {
static bool _registeredAll = false;
static Car create() {
if (!_registeredAll) { _registerAll()); }
/* ... */
}
}
// Inside of the same library, telsa_model_x.dart
class TelsaModelX extends Car {}
// Inside of the same library, global namespace:
// This method registers all "default" vehicles in the vehicle registery.
_registerAll() {
register('telsamodelx', () => new TelsaModelX());
}
// Inside of the same library, global namespace:
register(carName, carFxn) { /* ... */ }
Outside of the library, consumers had to call register(); somewhere to register their vehicle.
It is unnecessary duplication, and unfortunately separates the registration from the class in a way that makes it hard to track, but it's either cumbersome code or a performance hit by using dart:mirrors.
YMMV, but as the number of register-able items grow, I'm starting to look towards the dart:mirrors approach again.
I'm compiling three.d.ts (available from here) with the TypeScript develop branch. I get the following error:
Types of static property 'Utils' of class 'THREE.Shape' and class 'THREE.Path'
are incompatible
The problem is that
Shape defines a static Utils class
Shape indirectly inherits from Curve
Curve also defines a static Utils class with a signature unrelated to Shape.Utils
which is ill-formed according to the language spec. Summarised, three.d.ts contains something like the following code:
declare class A {
static Utils: {
f (): any;
}
}
declare class B extends A {
static Utils: {
// incompatible with A.Utils, without f(): any
g (): any;
}
}
Putting aside the question of why the type of a static member must be compatible with that of an inherited static member of the same name - which isn't the case in several other OO languages, but which does appear to be the case in TypeScript - I would like to know how to fix three.d.ts so I can compile it.
My current workaround is simply to copy and paste the signature of Curve.Utils into Shape.Utils, so that the latter structurally extends the former. But what is the "right" way to capture the signature of the underlying three.js file (here) in a .d.ts file? Is this a case of inheritance being used incorrectly?
Short answer is, per spec, Typescript doesn't allow member hiding through inheritance, like C# does automatically for example.
As defined in the language specifications section 8.2.3
A derived class inherits all members from its base class it doesn’t override. Inheritance means that a
derived class implicitly contains all non-overridden members of the base class. Both public and private
members are inherited, but only public members can be overridden. A member in a derived class is said
to override a member in a base class when the derived class member has the same name and kind
(instance or static) as the base class member. The type of an overriding member must be a subtype
(section 3.8.2) of the type of the overridden member, or otherwise a compile-time error occurs.
and
Base class static members can be overridden by derived class static members of any kind as >long as the
types are compatible, as described above.
Maybe they added some type checking into the latest compiler version which was missing before...
Here is a proposed solution, based on inheritance:
declare class UtilsA{
f():any;
}
declare class UtilsB extends UtilsA{
g():any;
}
declare class A {
static Utils:UtilsA;
}
declare class B extends A {
static Utils:UtilsB;
}
I am trying my best to explain the situation. I hope, what I wrote, is understandable.
We already have class defined like
public ref class TestClass
{
public:
TestClass();
virtual ~TestClass();
protected:
Car* m_car;
}
TestClass is managed C++ and Car is unmanaged C++.
So far so good, but now I need to make static object of TestClass also. So I modify the code like below
public ref class TestClass
{
private:
static TestClass^ s_test = nullptr ;
public:
TestClass();
virtual ~TestClass();
static TestClass^ Instance();
protected:
Car* m_car;
}
When I want to use static instant of the class, I just get it from calling
TestClass staticobj = TestClass::Instance();
Elsewhere, just call
TestClass normalobj = gcnew TestClass();
Instance function is creating s_test static object and returns it.
TestClass ^ TestClass::Instance()
{
if(s_test == nullptr)
{
s_test = gcnew TestClass();
s_test->m_car = new Car();
}
return s_test;
}
Is it a good approach?
Is there any other better approach to accomplish same thing?
Edit :
FYI Above code works.
I combined Krizz and Reed Copsey’s solutions. That solve independent Singleton and memory leak.
Here is my sample code,
Special Singleton class derived from test class,
public ref class SpecialSingletonTestClass: public TestClass
{
private:
static SpecialSingletonTestClass ^ s_ SpecialSingletonTestClass = nullptr;
public:
SpecialSingletonTestClass ();
static SpecialSingletonTestClass ^ Instance();
};
Changed the testclass so it has now one more finalizer function.
public ref class TestClass
{
public:
TestClass ();
virtual ~ TestClass ();
! TestClass ();
protected:
Car* m_car;
}
I tested above pattern , it worked.
Thanks you guys,
L.E.
Is it a good approach?
I would probably not consider this a good approach, as you're making a single class both a singleton and a normal class that you can instance directly.
Typically, if you need a singleton, this would preclude the need or desire to be able to instantiate the class.
If you truly need to have a way to have a "global" instance of this class, I would encapsulate that in a separate class which implements the singleton. This would, at least, make it clear that you are dealing with something that's a single instance in that case. I would not mix both use cases into a single class.
Well, actually there is an issue with memory leaks in your code.
You declare only virtual ~TestClass(); which, for managed classes, are internally turned by C++/CLI compiler into implementation of IDisposable.Dispose().
Therefore, if you put delete car into it, it will be called only if you delete test_class or, e.g. wrap into using (TestClass tst) {} block when using from C#.
It will not be called when object is GCed!
To be sure it is called you need to add finalizer to your class !MyClass(); which is turned by compiler into virtual void Finalize() and thus non-deterministically called when GC is freeing an object.
And it is the only way to free m_car of singleton object.
Therefore, I suggest:
TestClass()
{
m_car = new Car();
}
~TestClass()
{
if (m_car)
delete m_car;
m_car = NULL;
}
!TestClass()
{
if (m_car)
delete m_car;
m_car = NULL;
}
I'm unsure as to what situation you could possibly be in that would require both singleton-style semantics and normal creation semantics for the same class.
As far as what you've coded though, it looks completely fine. My only comments would be that your Instance() function shouldn't need to perform construction on Car, the Instance() function should just call the default constructor of TestClass which should do all that.
EDIT
In reference to:
#crush . The class is already define i just need to get static object of it. Singleton means only one object of the class, but in this case, class have multiple normal object. But i want to use only one object of this class for one specific goal only for limited period of time. – L.E. 2 mins ago
A singleton is (usually) a sign of bad design - alot of people call it an anti-pattern actually. Chances are if you just need this one single specific instance of this class for a limited period of time there are some issues:
Singleton design is made for static-style existence - the variable will live for the scope of your program after lazily initialized.
Allowing global access will move your code towards spaghetti logic. You'd be better off dynamically allocating the one you need and passing the pointer to it to where you need it to be. A shared_ptr would be good for this.
You should find a way around the singleton-style implementation in this case even if it's more work for you - it'll almost certainly be better design.
I would like to create a new array with a given type from a class object in GWT.
What I mean is I would like to emulate the functionality of
java.lang.reflect.Array.newInstance(Class<?> componentClass, int size)
The reason I need this to occur is that I have a library which occasionally needs to do the following:
Class<?> cls = array.getClass();
Class<?> cmp = cls.getComponentType();
This works if I pass it an array class normally, but I can't dynamically create a new array from some arbitrary component type.
I am well aware of GWT's lack of reflection; I understand this. However, this seems feasible even given GWT's limited reflection. The reason I believe this is that in the implementation, there exists an inaccessible static method for creating a class object for an array.
Similarly, I understand the array methods to just be type-safe wrappers around JavaScript arrays, and so should be easily hackable, even if JSNI is required.
In reality, the more important thing would be getting the class object, I can work around not being able to make new arrays.
If you are cool with creating a seed array of the correct type, you can use jsni along with some knowledge of super-super-source to create arrays WITHOUT copying through ArrayList (I avoid java.util overhead like the plague):
public static native <T> T[] newArray(T[] seed, int length)
/*-{
return #com.google.gwt.lang.Array::createFrom([Ljava/lang/Object;I)(seed, length);
}-*/;
Where seed is a zero-length array of the correct type you want, and length is the length you want (although, in production mode, arrays don't really have upper bounds, it makes the [].length field work correctly).
The com.google.gwt.lang package is a set of core utilities used in the compiler for base emulation, and can be found in gwt-dev!com/google/gwt/dev/jjs/intrinsic/com/google/gwt/lang.
You can only use these classes through jsni calls, and only in production gwt code (use if GWT.isProdMode()). In general, if you only access the com.google.gwt.lang classes in super-source code, you are guaranteed to never leak references to classes that only exist in compiled javascript.
if (GWT.isProdMode()){
return newArray(seed, length);
}else{
return Array.newInstance(seed.getComponentType(), length);
}
Note, you'll probably need to super-source the java.lang.reflect.Array class to avoid gwt compiler error, which suggests you'll want to put your native helper method there. However, I can't help you more than this, as it would overstep the bounds of my work contract.
The way that I did a similar thing was to pass an empty, 0 length array to the constructor of the object that will want to create the array from.
public class Foo extends Bar<Baz> {
public Foo()
{
super(new Baz[0]);
}
...
}
Baz:
public abstract class Baz<T>
{
private T[] emptyArray;
public Baz(T[] emptyArray)
{
this.emptyArray = emptyArray;
}
...
}
In this case the Bar class can't directly create new T[10], but we can do this:
ArrayList<T> al = new ArrayList<T>();
// add the items you want etc
T[] theArray = al.toArray(emptyArray);
And you get your array in a typesafe way (otherwise in your call super(new Baz[0]); will cause a compiler error).
I had to do something similar, I found it was possible using the Guava library's ObjectArrays class. Instead of the class object it requires a reference to an existing array.
T[] newArray = ObjectArrays.newArray(oldArray, oldArray.length);
For implementing an array concatenation method, I also stepped into the issue of missing Array.newInstance-method.
It's still not implemented, but if you have an existing array you can use
Arrays.copyOf(T[] original, int newLength)
instead.
I've got a CustomersModule.cs with the following Initialize() method:
public void Initialize()
{
container.RegisterType<ICustomersRepository, CustomersRepository>(new ContainerControlledLifetimeManager());
CustomersPresenter customersPresenter = this.container.Resolve<CustomersPresenter>();
}
The class I resolve from the container looks like this:
class CustomersPresenter
{
private CustomersView view;
private ICustomersRepository customersRespository;
public CustomersPresenter(CustomersView view,
ICustomersRepository customersRepository,
TestWhatever testWhatever)
{
this.view = view;
this.customersRespository = customersRepository;
}
}
The TestWhatever class is just a dummy class I created:
public class TestWhatever
{
public string Title { get; set; }
public TestWhatever()
{
Title = "this is the title";
}
}
Yet the container happily resolves CustomersPresenter even though I never registered it, and also the container somehow finds TestWhatever, instantiates it, and injects it into CustomersPresenter.
I was quite surprised to realize this since I couldn't find anywhere in the Prism documentation which explicitly stated that the container was so automatic.
So this is great, but it what else is the container doing that I don't know about i.e. what else can it do that I don't know about? For example, can I inject classes from other modules and if the modules happen to be loaded the container will inject them, and if not, it will inject a null?
There is nothing magical going on. You are specifying concrete types, so naturally they are resolvable, because if we have the Type object, we can call a constructor on it.
class Fred { };
Fred f1 = new Fred();
Type t = typeof(Fred);
Fred f2 = (Fred)t.GetConstructor(Type.EmptyTypes).Invoke(null);
The last line above is effectively what happens, the type t having been found by using typeof on the type parameter you give to Resolve.
If the type cannot be constructed by new (because it's in some unknown separate codebase) then you wouldn't be able to give it as a type parameter to Resolve.
In the second case, it is constructor injection, but it's still a known concrete constructable type. Via reflection, the Unity framework can get an array of all the Types of the parameters to the constructor. The type TestWhatever is constructable, so there is no ambiguity or difficulty over what to construct.
As to your concern about separate modules (assemblies), if you move TestWhatever to another assembly, that will not change the lines of code you've written; it will just mean that you have to add a reference to the other assembly to get this one to build. And then TestWhatever is still an unambiguously refeferenced constructable type, so it can be constructed by Unity.
In other words, if you can refer to the type in code, you can get a Type object, and so at runtime it will be directly constructable.
Response to comment:
If you delete the class TestWhatever, you will get a compile-time error, because you refer to that type in your code. So it won't be possible to get a runtime by doing that.
The decoupling is still in effect in this arrangement, because you could register a specific instance of TestWhatever, so every call to Resolve<TestWhatever>() will get the same instance, rather than constructing a new one.
The reason this works is because Unity is designed for it. When you Resolve with a concrete type, Unity looks to see if it can resolve from the container. If it cannot, then it just goes and instantiates the type resolving it's dependencies. It's really quite simple.