THIS PAGE IS UNDER CONSTRUCTION
Here I will talk about Halide
In order to learn Halide, let’s start with a simple example and see how Halide will produce the code for that example
Var x,y
Func grad;
Target target = get_host_target();
grad(x,y) = x + y;
Halide::Image<int32_t> output = gradient.realize(100, 200);
Now let’s see how the Halide will interpret this code. The first three lines are nothing but defining Var and Funcs and the target where we will generate the code. On line number 3, the update for the Func grad is written. This is where Halide creates and initializes the data-structures for the functional specification to be later compiled into a C/C++ code. Note that the program that you are writing is in itself in C++, and that is the beauty of Domain Specific Languages. You can crate an language from say C++ and then use that to generate efficient C++ code. So, as we all know the operator () has a highest precedence than = operator and both () and = operators of class Func have been overloaded. That means first the () operators on both LHS and RHS will be called and then the = operator will be called for the datatype returned by grad(x,y). As () operator with arguments as std::vector of Vars is overladed in Func.cpp and creates a FuncRef, the call to grad(x,y) returns a FuncRef object. As + has higher predence than =, so + will be evaluated on x and y, both of which are Vars and hence + operator. Now, if you search Var.cpp you will not find the overloaded operator + for Vars and the reason is there is a cast operator “operator Expr()” which automatically casts Var into object of type Expr and class Expr has overloaded + operator IROperator.h. This creates an Add IRNode. We will talk about IRNodes later, but the important thing is Add IRNode is also of derived from Expr and hence + on two Vars casted to Expr will return an Expr. Hence now, = operator will be called on FuncRef object with the argument as an Expr. The overloaded = operator for call FuncRef in Func.cpp which then calls func.define() or func.define_update() which is basically but creating a Defintion for the current functional expression. Each functional expression like the one line 4 has an associated definition with it. The first definition is called “initial definition” and the rest are “update definitions”. For example:
C(x,y) = 0.0f; // initial definition
C(x,y) = A(x,y) + B(x,y) // update definition
Also each Func object has a Function func data member, which is used for book keeping and is used in most of the places. Basically it is the internal representation of a Func. Hence func.define() in Function.cpp creates a Definition object for that particular function expression and copies it into the contents data member of the Function object. Then the call to = returns and we can execute the statement line 5 in our Halide code. At this point the the Halide program is in the memory and have not been compiled yet. realize() method will compile this program into a Halide IR which will then be lowered to LLVM IR or C.