Eigen
3.3.90 (mercurial changeset 94875feeeeb9)

Consider the following example program:
The goal of this page is to understand how Eigen compiles it, assuming that SSE2 vectorization is enabled (GCC option msse2).
Maybe you think, that the above example program is so simple, that compiling it shouldn't involve anything interesting. So before starting, let us explain what is nontrivial in compiling it correctly – that is, producing optimized code – so that the complexity of Eigen, that we'll explain here, is really useful.
Look at the line of code
The first important thing about compiling it, is that the arrays should be traversed only once, like
The problem is that if we make a naive C++ library where the VectorXf class has an operator+ returning a VectorXf, then the line of code (*) will amount to:
Obviously, the introduction of the temporary tmp here is useless. It has a very bad effect on performance, first because the creation of tmp requires a dynamic memory allocation in this context, and second as there are now two for loops:
Traversing the arrays twice instead of once is terrible for performance, as it means that we do many redundant memory accesses.
The second important thing about compiling the above program, is to make correct use of SSE2 instructions. Notice that Eigen also supports AltiVec and that all the discussion that we make here applies also to AltiVec.
SSE2, like AltiVec, is a set of instructions allowing to perform computations on packets of 128 bits at once. Since a float is 32 bits, this means that SSE2 instructions can handle 4 floats at once. This means that, if correctly used, they can make our computation go up to 4x faster.
However, in the above program, we have chosen size=50, so our vectors consist of 50 float's, and 50 is not a multiple of 4. This means that we cannot hope to do all of that computation using SSE2 instructions. The second best thing, to which we should aim, is to handle the 48 first coefficients with SSE2 instructions, since 48 is the biggest multiple of 4 below 50, and then handle separately, without SSE2, the 49th and 50th coefficients. Something like this:
So let us look line by line at our example program, and let's follow Eigen as it compiles it.
Let's analyze the first line:
First of all, VectorXf is the following typedef:
The class template Matrix is declared in src/Core/util/ForwardDeclarations.h with 6 template parameters, but the last 3 are automatically determined by the first 3. So you don't need to worry about them for now. Here, Matrix<float, Dynamic, 1> means a matrix of floats, with a dynamic number of rows and 1 column.
The Matrix class inherits a base class, MatrixBase. Don't worry about it, for now it suffices to say that MatrixBase is what unifies matrices/vectors and all the expressions types – more on that below.
When we do
the constructor that is called is Matrix::Matrix(int), in src/Core/Matrix.h. Besides some assertions, all it does is to construct the m_storage member, which is of type DenseStorage<float, Dynamic, Dynamic, 1>.
You may wonder, isn't it overengineering to have the storage in a separate class? The reason is that the Matrix class template covers all kinds of matrices and vector: both fixedsize and dynamicsize. The storage method is not the same in these two cases. For fixedsize, the matrix coefficients are stored as a plain member array. For dynamicsize, the coefficients will be stored as a pointer to a dynamicallyallocated array. Because of this, we need to abstract storage away from the Matrix class. That's DenseStorage.
Let's look at this constructor, in src/Core/DenseStorage.h. You can see that there are many partial template specializations of DenseStorages here, treating separately the cases where dimensions are Dynamic or fixed at compiletime. The partial specialization that we are looking at is:
Here, the constructor called is DenseStorage::DenseStorage(int size, int rows, int columns) with size=50, rows=50, columns=1.
Here is this constructor:
Here, the m_data member is the actual array of coefficients of the matrix. As you see, it is dynamically allocated. Rather than calling new[] or malloc(), as you can see, we have our own internal::aligned_new defined in src/Core/util/Memory.h. What it does is that if vectorization is enabled, then it uses a platformspecific call to allocate a 128bitaligned array, as that is very useful for vectorization with both SSE2 and AltiVec. If vectorization is disabled, it amounts to the standard new[].
As you can see, the constructor also sets the m_rows member to size. Notice that there is no m_columns member: indeed, in this partial specialization of DenseStorage, we know the number of columns at compiletime, since the _Cols template parameter is different from Dynamic. Namely, in our case, _Cols is 1, which is to say that our vector is just a matrix with 1 column. Hence, there is no need to store the number of columns as a runtime variable.
When you call VectorXf::data() to get the pointer to the array of coefficients, it returns DenseStorage::data() which returns the m_data member.
When you call VectorXf::size() to get the size of the vector, this is actually a method in the base class MatrixBase. It determines that the vector is a columnvector, since ColsAtCompileTime==1 (this comes from the template parameters in the typedef VectorXf). It deduces that the size is the number of rows, so it returns VectorXf::rows(), which returns DenseStorage::rows(), which returns the m_rows member, which was set to size by the constructor.
Now that our vectors are constructed, let's move on to the next line:
The executive summary is that operator+ returns a "sum of vectors" expression, but doesn't actually perform the computation. It is the operator=, whose call occurs thereafter, that does the computation.
Let us now see what Eigen does when it sees this:
Here, v and w are of type VectorXf, which is a typedef for a specialization of Matrix (as we explained above), which is a subclass of MatrixBase. So what is being called is
The return type of this operator is
The CwiseBinaryOp class is our first encounter with an expression template. As we said, the operator+ doesn't by itself perform any computation, it just returns an abstract "sum of vectors" expression. Since there are also "difference of vectors" and "coefficientwise product of vectors" expressions, we unify them all as "coefficientwise binary operations", which we abbreviate as "CwiseBinaryOp". "Coefficientwise" means that the operations is performed coefficient by coefficient. "binary" means that there are two operands – we are adding two vectors with one another.
Now you might ask, what if we did something like
The first v + w would return a CwiseBinaryOp as above, so in order for this to compile, we'd need to define an operator+ also in the class CwiseBinaryOp... at this point it starts looking like a nightmare: are we going to have to define all operators in each of the expression classes (as you guessed, CwiseBinaryOp is only one of many) ? This looks like a dead end!
The solution is that CwiseBinaryOp itself, as well as Matrix and all the other expression types, is a subclass of MatrixBase. So it is enough to define once and for all the operators in class MatrixBase.
Since MatrixBase is the common base class of different subclasses, the aspects that depend on the subclass must be abstracted from MatrixBase. This is called polymorphism.
The classical approach to polymorphism in C++ is by means of virtual functions. This is dynamic polymorphism. Here we don't want dynamic polymorphism because the whole design of Eigen is based around the assumption that all the complexity, all the abstraction, gets resolved at compiletime. This is crucial: if the abstraction can't get resolved at compiletime, Eigen's compiletime optimization mechanisms become useless, not to mention that if that abstraction has to be resolved at runtime it'll incur an overhead by itself.
Here, what we want is to have a single class MatrixBase as the base of many subclasses, in such a way that each MatrixBase object (be it a matrix, or vector, or any kind of expression) knows at compiletime (as opposed to runtime) of which particular subclass it is an object (i.e. whether it is a matrix, or an expression, and what kind of expression).
The solution is the Curiously Recurring Template Pattern. Let's do the break now. Hopefully you can read this wikipedia page during the break if needed, but it won't be allowed during the exam.
In short, MatrixBase takes a template parameter Derived. Whenever we define a subclass Subclass, we actually make Subclass inherit MatrixBase<Subclass>. The point is that different subclasses inherit different MatrixBase types. Thanks to this, whenever we have an object of a subclass, and we call on it some MatrixBase method, we still remember even from inside the MatrixBase method which particular subclass we're talking about.
This means that we can put almost all the methods and operators in the base class MatrixBase, and have only the bare minimum in the subclasses. If you look at the subclasses in Eigen, like for instance the CwiseBinaryOp class, they have very few methods. There are coeff() and sometimes coeffRef() methods for access to the coefficients, there are rows() and cols() methods returning the number of rows and columns, but there isn't much more than that. All the meat is in MatrixBase, so it only needs to be coded once for all kinds of expressions, matrices, and vectors.
So let's end this digression and come back to the piece of code from our example program that we were currently analyzing,
Now that MatrixBase is a good friend, let's write fully the prototype of the operator+ that gets called here (this code is from src/Core/MatrixBase.h):
Here of course, Derived and OtherDerived are VectorXf.
As we said, CwiseBinaryOp is also used for other operations such as substration, so it takes another template parameter determining the operation that will be applied to coefficients. This template parameter is a functor, that is, a class in which we have an operator() so it behaves like a function. Here, the functor used is internal::scalar_sum_op. It is defined in src/Core/Functors.h.
Let us now explain the internal::traits here. The internal::scalar_sum_op class takes one template parameter: the type of the numbers to handle. Here of course we want to pass the scalar type (a.k.a. numeric type) of VectorXf, which is float
. How do we determine which is the scalar type of Derived ? Throughout Eigen, all matrix and expression types define a typedef Scalar which gives its scalar type. For example, VectorXf::Scalar is a typedef for float
. So here, if life was easy, we could find the numeric type of Derived as just
Unfortunately, we can't do that here, as the compiler would complain that the type Derived hasn't yet been defined. So we use a workaround: in src/Core/util/ForwardDeclarations.h, we declared (not defined!) all our subclasses, like Matrix, and we also declared the following class template:
In src/Core/Matrix.h, right before the definition of class Matrix, we define a partial specialization of internal::traits for T=Matrix<any template parameters>. In this specialization of internal::traits, we define the Scalar typedef. So when we actually define Matrix, it is legal to refer to "typename internal::traits\<Matrix\>::Scalar".
Anyway, we have declared our operator+. In our case, where Derived and OtherDerived are VectorXf, the above declaration amounts to:
Let's now jump to src/Core/CwiseBinaryOp.h to see how it is defined. As you can see there, all it does is to return a CwiseBinaryOp object, and this object is just storing references to the lefthandside and righthandside expressions – here, these are the vectors v and w. Well, the CwiseBinaryOp object is also storing an instance of the (empty) functor class, but you shouldn't worry about it as that is a minor implementation detail.
Thus, the operator+ hasn't performed any actual computation. To summarize, the operation v + w just returned an object of type CwiseBinaryOp which did nothing else than just storing references to v and w.
At this point, the expression v + w has finished evaluating, so, in the process of compiling the line of code
we now enter the operator=.
What operator= is being called here? The vector u is an object of class VectorXf, i.e. Matrix. In src/Core/Matrix.h, inside the definition of class Matrix, we see this:
Here, Base is a typedef for MatrixBase<Matrix>. So, what is being called is the operator= of MatrixBase. Let's see its prototype in src/Core/MatrixBase.h:
Here, Derived is VectorXf (since u is a VectorXf) and OtherDerived is CwiseBinaryOp. More specifically, as explained in the previous section, OtherDerived is:
So the full prototype of the operator= being called is:
This operator= literally reads "copying a sum of two VectorXf's into another VectorXf".
Let's now look at the implementation of this operator=. It resides in the file src/Core/Assign.h.
What we can see there is:
OK so our next task is to understand internal::assign_selector :)
Here is its declaration (all that is still in the same file src/Core/Assign.h)
So internal::assign_selector takes 4 template parameters, but the 2 last ones are automatically determined by the 2 first ones.
EvalBeforeAssigning is here to enforce the EvalBeforeAssigningBit. As explained here, certain expressions have this flag which makes them automatically evaluate into temporaries before assigning them to another expression. This is the case of the Product expression, in order to avoid strange aliasing effects when doing "m = m * m;" However, of course here our CwiseBinaryOp expression doesn't have the EvalBeforeAssigningBit: we said since the beginning that we didn't want a temporary to be introduced here. So if you go to src/Core/CwiseBinaryOp.h, you'll see that the Flags in internal::traits<CwiseBinaryOp> don't include the EvalBeforeAssigningBit. The Flags member of CwiseBinaryOp is then imported from the internal::traits by the EIGEN_GENERIC_PUBLIC_INTERFACE macro. Anyway, here the template parameter EvalBeforeAssigning has the value false
.
NeedToTranspose is here for the case where the user wants to copy a rowvector into a columnvector. We allow this as a special exception to the general rule that in assignments we require the dimesions to match. Anyway, here both the lefthand and righthand sides are column vectors, in the sense that ColsAtCompileTime is equal to 1. So NeedToTranspose is false
too.
So, here we are in the partial specialization:
Here's how it is defined:
OK so now our next job is to understand how lazyAssign works :)
What do we see here? Some assertions, and then the only interesting line is:
OK so now we want to know what is inside internal::assign_impl.
Here is its declaration:
Again, internal::assign_selector takes 4 template parameters, but the 2 last ones are automatically determined by the 2 first ones.
These two parameters Vectorization and Unrolling are determined by a helper class internal::assign_traits. Its job is to determine which vectorization strategy to use (that is Vectorization) and which unrolling strategy to use (that is Unrolling).
We'll not enter into the details of how these strategies are chosen (this is in the implementation of internal::assign_traits at the top of the same file). Let's just say that here Vectorization has the value LinearVectorization, and Unrolling has the value NoUnrolling (the latter is obvious since our vectors have dynamic size so there's no way to unroll the loop at compiletime).
So the partial specialization of internal::assign_impl that we're looking at is:
Here is how it's defined:
Here's how it works. LinearVectorization means that the lefthand and righthand side expression can be accessed linearly i.e. you can refer to their coefficients by one integer index, as opposed to having to refer to its coefficients by two integers row, column.
As we said at the beginning, vectorization works with blocks of 4 floats. Here, PacketSize is 4.
There are two potential problems that we need to deal with:
Now come the actual loops.
First, the vectorized part: the 48 first coefficients out of 50 will be copied by packets of 4:
What is copyPacket? It is defined in src/Core/Coeffs.h:
OK, what are writePacket() and packet() here?
First, writePacket() here is a method on the lefthand side VectorXf. So we go to src/Core/Matrix.h to look at its definition:
Here, StoreMode is Aligned, indicating that we are doing a 128bitaligned write access, PacketScalar is a type representing a "SSE packet of 4 floats" and internal::pstoret is a function writing such a packet in memory. Their definitions are architecturespecific, we find them in src/Core/arch/SSE/PacketMath.h:
The line in src/Core/arch/SSE/PacketMath.h that determines the PacketScalar type (via a typedef in Matrix.h) is:
Here, __m128 is a SSEspecific type. Notice that the enum size here is what was used to define packetSize above.
And here is the implementation of internal::pstoret:
Here, __mm_store_ps is a SSEspecific intrinsic function, representing a single SSE instruction. The difference between internal::pstore and internal::pstoret is that internal::pstoret is a dispatcher handling both the aligned and unaligned cases, you find its definition in src/Core/GenericPacketMath.h:
OK, that explains how writePacket() works. Now let's look into the packet() call. Remember that we are analyzing this line of code inside copyPacket():
Here, other is our sum expression v + w. The .derived() is just casting from MatrixBase to the subclass which here is CwiseBinaryOp. So let's go to src/Core/CwiseBinaryOp.h:
Here, m_lhs is the vector v, and m_rhs is the vector w. So the packet() function here is Matrix::packet(). The template parameter LoadMode is Aligned. So we're looking at
We let you look up the definition of internal::ploadt in GenericPacketMath.h and the internal::pload in src/Core/arch/SSE/PacketMath.h. It is very similar to the above for internal::pstore.
Let's go back to CwiseBinaryOp::packet(). Once the packets from the vectors v and w have been returned, what does this function do? It calls m_functor.packetOp() on them. What is m_functor? Here we must remember what particular template specialization of CwiseBinaryOp we're dealing with:
So m_functor is an object of the empty class internal::scalar_sum_op<float>. As we mentioned above, don't worry about why we constructed an object of this empty class at all – it's an implementation detail, the point is that some other functors need to store member data.
Anyway, internal::scalar_sum_op is defined in src/Core/Functors.h:
As you can see, all what packetOp() does is to call internal::padd on the two packets. Here is the definition of internal::padd from src/Core/arch/SSE/PacketMath.h:
Here, _mm_add_ps is a SSEspecific intrinsic function, representing a single SSE instruction.
To summarize, the loop
has been compiled to the following code: for index going from 0 to the 11 ( = 48/4  1), read the ith packet (of 4 floats) from the vector v and the ith packet from the vector w using two __mm_load_ps SSE instructions, then add them together using a __mm_add_ps instruction, then store the result using a __mm_store_ps instruction.
There remains the second loop handling the last few (here, the last 2) coefficients:
However, it works just like the one we just explained, it is just simpler because there is no SSE vectorization involved here. copyPacket() becomes copyCoeff(), packet() becomes coeff(), writePacket() becomes coeffRef(). If you followed us this far, you can probably understand this part by yourself.
We see that all the C++ abstraction of Eigen goes away during compilation and that we indeed are precisely controlling which assembly instructions we emit. Such is the beauty of C++! Since we have such precise control over the emitted assembly instructions, but such complex logic to choose the right instructions, we can say that Eigen really behaves like an optimizing compiler. If you prefer, you could say that Eigen behaves like a script for the compiler. In a sense, C++ template metaprogramming is scripting the compiler – and it's been shown that this scripting language is Turingcomplete. See Wikipedia.