Hybrid Inter-function call wrong

#1

Hello, everyone, I’m a fresh man of tvm, recently i run the test code about hybrid in “tests/python/unittest/test_hybrid_script.py”。When i run test_func_call(), everything is OK,but when i change the called func name, there has something wrong, I can’t find the reason.Can you help me?
Thank you very much.
My code is:

def test_func_call():
    @tvm.hybrid.script
    def outer_productss(n, m, a, b):
        """This is a simple outer product."""
        c = output_tensor((n,m), a.dtype)
        for i in range(n):
            for j in range(m):
                assert i<n and j<m, "index out of range!"
                c[i,j] = a[i]*b[j]
        return c

    @tvm.hybrid.script
    def foo(a, b):
        for i in range(len(a)):
            a[i] = i + 1.0
        for i in range(len(a)):
            b[i] = i + 1.0
        c = outer_productss(10, 10, a, b)
        d = output_tensor(c.shape, c.dtype)
        for i in range(10):
            for j in range(10):
                d[i, j] = c[i, j] + i * j
        return d

    a = tvm.placeholder((10, ), name='a')
    b = tvm.placeholder((10, ), name='b')
    func, ins, outs = run_and_check(foo, [a, b])
    run_and_check(func, ins, outs=outs)

Error  is : 
Traceback (most recent call last):
  File "test_hybrid.py", line 380, in <module>
    test_func_call()
  File "test_hybrid.py", line 332, in test_func_call
    func, ins, outs = run_and_check(foo, [a, b])
  File "test_hybrid.py", line 22, in run_and_check
    outs = func(*tuple(tvm.convert(i) if isinstance(i,list) else i for i in args))
  File "</home/s50004646/.local/lib/python3.5/site-packages/decorator.py:decorator-gen-4>", line 2, in foo
  File "/home/s50004646/tvm-master/python/tvm/hybrid/__init__.py", line 62, in wrapped_func
    return source_to_op(src, args, func.__globals__, closure_vars)
  File "/home/s50004646/tvm-master/python/tvm/hybrid/parser.py", line 652, in source_to_op
    parser = parse_python(src, args, symbols, closure_vars)
  File "/home/s50004646/tvm-master/python/tvm/hybrid/parser.py", line 620, in parse_python
    var_usage = determine_variable_usage(root, args, symbols, closure_vars)
  File "/home/s50004646/tvm-master/python/tvm/hybrid/preprocessor.py", line 116, in determine_variable_usage
    visitor.visit(root)
  File "/usr/lib/python3.5/ast.py", line 245, in visit
    return visitor(node)
  File "/usr/lib/python3.5/ast.py", line 253, in generic_visit
    self.visit(item)
  File "/usr/lib/python3.5/ast.py", line 245, in visit
    return visitor(node)
  File "/home/s50004646/tvm-master/python/tvm/hybrid/preprocessor.py", line 46, in visit_FunctionDef
    self.visit(i)
  File "/usr/lib/python3.5/ast.py", line 245, in visit
    return visitor(node)
  File "/usr/lib/python3.5/ast.py", line 255, in generic_visit
    self.visit(value)
  File "/usr/lib/python3.5/ast.py", line 245, in visit
    return visitor(node)
  File "/home/s50004646/tvm-master/python/tvm/hybrid/preprocessor.py", line 66, in visit_Call
    "Function call id not in intrinsics' list")
  File "/home/s50004646/tvm-master/python/tvm/hybrid/util.py", line 42, in _internal_assert
    raise ValueError(err)
ValueError: Function call id not in intrinsics' list
#2

CC: @were if you have bandwidth

#3

This is a bug. I apologize.
If you move productss as a global function, you can work this around.
I will be fixing it as soon as possible.

#4

@were as you said,I move productss as a global function, it works, Thank you very much.

#5

@junrushao can TensorIR support function call now? I mean the function decorated by @tvm.script.tir call another function.

#6

No it’s not supported yet, you can write the callee function as a inner block inside the caller block.

#7

will you have plan to support it?

#8

Hi @chengchen , would you mind elaborating more on the benefits of supporting this feature? (e.g. in what case inter-function call would simplify op definition)?

#10

i hope to provide builtin tensor operation based on tir like tvm/python/tvm/topi/math.py. In buitin tensor operation, it can be programed as tir script as well and therefore can be scheduled.