Hybrid Frontend Language Reference¶
This hybrid frontend allows users to write preliminary versions of some idioms that yet have been supported by TVM officially.
Both software emulation and compilation are supported. To define a function,
you need to use
tvm.hybrid.script decorator to indicate this is a hybrid function:
@tvm.hybrid.script def outer_product(a, b, c): c = output_tensor((100, 99), 'float32') for i in range(a.shape): for j in range(b.shape): c[i, j] = a[i] * b[j] return c a = numpy.random.randn(100) b = numpy.random.randn(99) c = outer_product(a, b)
This decorator will import Keywords required spontaneously when software emulation. After software emulation is done, the imported keywords will be cleaned up. Users do not need worry about keyword conflict and pollution.
Every element passed for software emulation in the argument list is either a python variable
numpy numeric type.
This function is not encouraged to use, users are encouraged to use the second interface. The current parse interface looks like:
a = tvm.placeholder((100, ), name='a') b = tvm.placeholder((99, ), name='b') parser = tvm.hybrid.parse(outer_product, [a, b]) # return the parser of this function
If we pass these tvm data structures, like
tvm.container.Array, to this function, it returns a op node:
a = tvm.placeholder((100, ), name='a') b = tvm.placeholder((99, ), name='b') c = outer_product(a, b, c) # return the output tensor(s) of the operator
You can use any methods that can be applied on a TVM
OpNode, like create_schedule, although
so far, the functionality of schedule is as limited as
ExternOpNode. At least, it can be built
to LLVM module.
Follow up the example above, you can use some tvm like interfaces to tune the code:
i, j = c.op.axis sch = tvm.create_schedule(op) jo, ji = sch.split(j, 4) sch.vectorize(ji)
For now, you can use loop annotations (
loop manipulation (
This is a preliminary function, so users should be in charge of the correctness of the functionality after tuning. Specifically, users should be careful when fusing and reorderding imperfect loops.
In HalideIR, loops have in total 4 types:
Here we use
these 4 keywords to annotate the corresponding types of for loops.
The the usage is roughly the same as Python standard
Besides all the loop types supported in Halide,
const_range is supported for some specific conditions.
tvm.container.Array is desired to pass as an argument, but in TVM-HalideIR, there is no
such support that converts
tvm.container.Array to an
Expr. Thus, a limited feature is supported.
Users can access containers by either constants or constants loops annotated.
@tvm.hybrid.script def foo(a, b): # b is a tvm.container.Array c = output_tensor(a.shape, a.dtype) for i in const_range(len(a)): # because you have b access, i should be explicitly annotated as const_range c[i] = a[i] + b[i] return c
All the mutatable variables will be lowered to an array with size 1. It regards the first store of a variable as its declaration.
Unlike conventional Python, in hybrid script, the declared variable can only be used in the scope level it is declared.
Currently, you can ONLY use basic-typed variables, i.e. the type of the
variable should be either
for i in range(5): s = 0 # declaration, this s will be a 1-array in lowered IR for j in range(5): s += a[i, j] # do something with sum b[i] = sum # you can still use sum in this level a = s # you CANNOT use s here, even though it is allowed in conventional Python
So far, ONLY tensors’
dtype attribute are supported!
shape atrribute is essentailly a tuple, so you MUST access it as an array.
Currently, only constant-indexed access is supported.
x = a.shape # OK! for i in range(3): for j in a.shape[i]: # BAD! i is not a constant! # do something
Conditional Statement and Expression¶
if condition1 and condition2 and condition3: # do something else: # do something else # Select a = b if condition else c
False keyword supported yet.
So far, these math intrinsics,
popcount, are supported.
No import is required, just as it is mentioned in Software Emulation, just use it!
Under construction, this function will be supported later!
Use a function call
allocation(shape, type, share/local) to declare an array buffer.
The basic usage is roughly the same as a normal
numpy.array, and you should access
high-dim array in
a[i, j, k] fashion instead of
tvm.container.Array for compilation.
You can also do loop-thread bind by writing code like this:
for tx in bind("threadIdx.x", 100): a[tx] = b[tx]
Assert statement is supported, you can simply use it as it is in standard Python.
assert cond, mesg
Assert is NOT a function call. Users are encouraged to use assert in the way
presented above — condition followed by message. It fits both Python AST and HalideIR.
- For keywords:
- Math keywords: