Hybrid Frontend Developer Guide¶
If you are a developer:
1. who is trying writing some preliminary patterns that have not been supported by TVM yet, maybe Hybrid Frontend Language Reference is a better place for you.
who wants to know the implementation details of this module, you are right here!
In software emulation, the most interesting thing is the decorator
This decorator helps 2 things:
Importing runtime variables
Overloading the function according to the arguments passed
Correct me if I am wrong: I believe that how 1. is implemented is dangerous, but I have no
choice. What I did is to add those names into python dict
func.__global__ and after
the call to
func is done, those names will be cleaned up.
Overload is simple: the decorator checks the arguments’ types and determines which function should be actually called.
Compilation is a large module, you can see
python/tvm/te/hybrid/ for more
details. The first stage determines the usage, or more accurately the
declaration of each variable and the second stage does the actual IR
So far, ONLY tensors’ shape attribute is supported. You can see
python/tvm/te/hybrid/parser.py for more details. This is a hacky solution, I just
check the attributes when subscript.
In HalideIR, loops have in total 4 types:
Unlike what that is in HalideIR, in
a is the starting point and
is the trip count of iterations. Here
loop_type(a, b) indicates
[a, b). Thus, when lowering it
to HalideIR, we need to do
start, extent = a, b - a
In HalideIR those are enums, they are in passive form. Here we use active form to annotate loops, because they are ready to run.
Because there is no variables in
HalideIR, all the mutable variables will be lowered to an array with size 1.
It takes the first store of a variable as its declaration.
So far, these math intrinsics,
popcount, are supported.
Math intrinsics will be imported by the decorator. Most of the intrinsics are borrowed by library implementation
sigmoid. I implemented them manually.
You can cast values by using the keywords