Jitclass + numpy functions can't compile with attribute as second parameter

[JulVandenBroeck]

Hello everyone!

Thanks for the great work on numba.

I believe I have encountered a bug that happens when the second argument of numpy.broadcast_to is an object attribute.

Below is a MWE that triggers the bug (notice the self.n as the second argument of np.broadcast_to):

import numpy as np
from numba import int32, float64
from numba.experimental import jitclass

@jitclass([
    ("n" , int32     ),
    ("a" , float64[:]),
])
class Foo:
    def __init__(self, a, n):
        self.n = n
        self.a = np.broadcast_to([a], self.n)

Foo(1., 6)

Running this code raises the following exceptions:

Traceback (most recent call last):
  File "c:\Users\julvd\repos\transport\wigner\NpNNp\test.py", line 12, in <module>
    Test(1., 6)
  File "c:\Users\julvd\repos\transport\.venv\Lib\site-packages\numba\experimental\jitclass\base.py", line 124, in __call__
    return cls._ctor(*bind.args[1:], **bind.kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "c:\Users\julvd\repos\transport\.venv\Lib\site-packages\numba\core\dispatcher.py", line 468, in _compile_for_args
    error_rewrite(e, 'typing')
  File "c:\Users\julvd\repos\transport\.venv\Lib\site-packages\numba\core\dispatcher.py", line 409, in error_rewrite
    raise e.with_traceback(None)
numba.core.errors.TypingError: Failed in nopython mode pipeline (step: nopython frontend)
Failed in nopython mode pipeline (step: nopython frontend)
No implementation of function Function(<function broadcast_to at 0x000001737F643D80>) found for signature:
 
 >>> broadcast_to(list(float64)<iv=None>, int32)
 
There are 2 candidate implementations:
      - Of which 2 did not match due to:
      Overload in function 'numpy_broadcast_to': File: numba\np\arrayobj.py: Line 1461.
        With argument(s): '(list(float64)<iv=None>, int32)':
       Rejected as the implementation raised a specific error:
         TypingError: Failed in nopython mode pipeline (step: nopython frontend)
       No implementation of function Function(<function broadcast_to at 0x000001737F643D80>) found for signature:
        
        >>> broadcast_to(list(float64)<iv=None>, UniTuple(int32 x 1))
        
       There are 2 candidate implementations:
             - Of which 2 did not match due to:
             Overload in function 'numpy_broadcast_to': File: numba\np\arrayobj.py: Line 1461.
               With argument(s): '(list(float64)<iv=None>, UniTuple(int32 x 1))':
              Rejected as the implementation raised a specific error:
                AssertionError: Failed in nopython mode pipeline (step: native lowering)
       
         raised from c:\Users\julvd\repos\transport\.venv\Lib\site-packages\llvmlite\ir\instructions.py:317

       During: resolving callee type: Function(<function broadcast_to at 0x000001737F643D80>)
       During: typing of call at c:\Users\julvd\repos\transport\.venv\Lib\site-packages\numba\np\arrayobj.py (1469)
       

       File ".venv\Lib\site-packages\numba\np\arrayobj.py", line 1469:
               def impl(array, shape):
                   return np.broadcast_to(array, (shape,))
                   ^

  raised from c:\Users\julvd\repos\transport\.venv\Lib\site-packages\numba\core\typeinfer.py:1091

During: resolving callee type: Function(<function broadcast_to at 0x000001737F643D80>)
During: typing of call at c:\Users\julvd\repos\transport\wigner\NpNNp\test.py (11)


File "wigner\NpNNp\test.py", line 11:
    def __init__(self, a, n):
        <source elided>
        self.n = n
        self.a = np.broadcast_to([a], self.n)
        ^

During: resolving callee type: jitclass.Test#17301c50ec0<n:int32,a:array(float64, 1d, A)>
During: typing of call at <string> (3)

During: resolving callee type: jitclass.Test#17301c50ec0<n:int32,a:array(float64, 1d, A)>
During: typing of call at <string> (3)


File "<string>", line 3:
<source missing, REPL/exec in use?>


(.venv) C:\Users\julvd\repos\transport>c:\Users\julvd\repos\transport\.venv\Scripts\python.exe "c:\Users\julvd\repos\transport\mwe.py"
Traceback (most recent call last):
  File "c:\Users\julvd\repos\transport\mwe.py", line 12, in <module>
    Foo(1., 6)
  File "c:\Users\julvd\repos\transport\.venv\Lib\site-packages\numba\experimental\jitclass\base.py", line 124, in __call__
    return cls._ctor(*bind.args[1:], **bind.kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "c:\Users\julvd\repos\transport\.venv\Lib\site-packages\numba\core\dispatcher.py", line 468, in _compile_for_args
    error_rewrite(e, 'typing')
  File "c:\Users\julvd\repos\transport\.venv\Lib\site-packages\numba\core\dispatcher.py", line 409, in error_rewrite
    raise e.with_traceback(None)
numba.core.errors.TypingError: Failed in nopython mode pipeline (step: nopython frontend)
Failed in nopython mode pipeline (step: nopython frontend)
No implementation of function Function(<function broadcast_to at 0x000002A60B9C3D80>) found for signature:
 
 >>> broadcast_to(list(float64)<iv=None>, int32)
 
There are 2 candidate implementations:
      - Of which 2 did not match due to:
      Overload in function 'numpy_broadcast_to': File: numba\np\arrayobj.py: Line 1461.
        With argument(s): '(list(float64)<iv=None>, int32)':
       Rejected as the implementation raised a specific error:
         TypingError: Failed in nopython mode pipeline (step: nopython frontend)
       No implementation of function Function(<function broadcast_to at 0x000002A60B9C3D80>) found for signature:
        
        >>> broadcast_to(list(float64)<iv=None>, UniTuple(int32 x 1))
        
       There are 2 candidate implementations:
             - Of which 2 did not match due to:
             Overload in function 'numpy_broadcast_to': File: numba\np\arrayobj.py: Line 1461.
               With argument(s): '(list(float64)<iv=None>, UniTuple(int32 x 1))':
              Rejected as the implementation raised a specific error:
                AssertionError: Failed in nopython mode pipeline (step: native lowering)
       
         raised from c:\Users\julvd\repos\transport\.venv\Lib\site-packages\llvmlite\ir\instructions.py:317

       During: resolving callee type: Function(<function broadcast_to at 0x000002A60B9C3D80>)
       During: typing of call at c:\Users\julvd\repos\transport\.venv\Lib\site-packages\numba\np\arrayobj.py (1469)
       

       File ".venv\Lib\site-packages\numba\np\arrayobj.py", line 1469:
               def impl(array, shape):
                   return np.broadcast_to(array, (shape,))
                   ^

  raised from c:\Users\julvd\repos\transport\.venv\Lib\site-packages\numba\core\typeinfer.py:1091

During: resolving callee type: Function(<function broadcast_to at 0x000002A60B9C3D80>)
During: typing of call at c:\Users\julvd\repos\transport\mwe.py (11)


File "mwe.py", line 11:
    def __init__(self, a, n):
        <source elided>
        self.n = n
        self.a = np.broadcast_to([a], self.n)
        ^

During: resolving callee type: jitclass.Foo#2a60bc9df10<n:int32,a:array(float64, 1d, A)>
During: typing of call at <string> (3)

During: resolving callee type: jitclass.Foo#2a60bc9df10<n:int32,a:array(float64, 1d, A)>
During: typing of call at <string> (3)


File "<string>", line 3:
<source missing, REPL/exec in use?>

In contrast, using the constructor argument n itself gives no errors:

import numpy as np
from numba import int32, float64
from numba.experimental import jitclass
@jitclass([
    ("n" , int32     ),
    ("a" , float64[:]),
])
class Foo:
    def __init__(self, a, n):
        self.n = n
        self.a = np.broadcast_to([a], n)
Foo(1., 6)

Thanks for taking the time.

Kind regards
Jul

[JulVandenBroeck]

I just noticed that a similar thing happens for np.trapz.

MWE:

import numpy as np
from numba import float64
from numba.experimental import jitclass
@jitclass([
    ("xx" , float64[:]),
    ("res", float64   ),
])
class Foo:
    def __init__(self, aa, xx):
        self.xx = xx
        self.res = np.trapz(aa, x=self.xx)
aa = np.asarray([1,4,9,16,25], dtype=np.float_)
xx = np.asarray([1,2,3,4,5], dtype=np.float_)
print(Foo(aa, xx).res)

This throws the following error:

Traceback (most recent call last):
  File "c:\Users\julvd\repos\transport\mwe.py", line 14, in <module>
    print(Foo(aa, xx).res)
          ^^^^^^^^^^^
  File "c:\Users\julvd\repos\transport\.venv\Lib\site-packages\numba\experimental\jitclass\base.py", line 124, in __call__
    return cls._ctor(*bind.args[1:], **bind.kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "c:\Users\julvd\repos\transport\.venv\Lib\site-packages\numba\core\dispatcher.py", line 468, in _compile_for_args
    error_rewrite(e, 'typing')
  File "c:\Users\julvd\repos\transport\.venv\Lib\site-packages\numba\core\dispatcher.py", line 409, in error_rewrite
    raise e.with_traceback(None)
numba.core.errors.TypingError: Failed in nopython mode pipeline (step: nopython frontend)
Failed in nopython mode pipeline (step: nopython frontend)
No implementation of function Function(<function trapz at 0x000001D97FECC720>) found for signature:
 
 >>> trapz(array(float64, 1d, C), x=array(float64, 1d, A))
 
There are 2 candidate implementations:
  - Of which 2 did not match due to:
  Overload in function 'np_trapz': File: numba\np\arraymath.py: Line 2214.
    With argument(s): '(array(float64, 1d, C), x=array(float64, 1d, A))':
   Rejected as the implementation raised a specific error:
     TypingError: Failed in nopython mode pipeline (step: nopython frontend)
   No implementation of function Function(<function _get_d at 0x000001D924360400>) found for signature:
    
    >>> _get_d(array(float64, 1d, A), float64)
    
   There are 2 candidate implementations:
         - Of which 2 did not match due to:
         Overload in function 'get_d_impl': File: numba\np\arraymath.py: Line 2203.
           With argument(s): '(array(float64, 1d, A), float64)':
          Rejected as the implementation raised a specific error:
            TypingError: Failed in nopython mode pipeline (step: nopython frontend)
          No implementation of function Function(<function diff at 0x000001D97FEC5A80>) found for signature:

           >>> diff(array(float64, 1d, A))

          There are 2 candidate implementations:
                - Of which 2 did not match due to:
                Overload in function 'np_diff_impl': File: numba\np\arraymath.py: Line 3525.
                  With argument(s): '(array(float64, 1d, A))':
                 Rejected as the implementation raised a specific error:
                   TypingError: Failed in nopython mode pipeline (step: nopython frontend)
                 - Resolution failure for literal arguments:
                 reshape() supports contiguous array only
                 - Resolution failure for non-literal arguments:
                 reshape() supports contiguous array only

                 During: resolving callee type: BoundFunction(array.reshape for array(float64, 1d, A))
                 During: typing of call at c:\Users\julvd\repos\transport\.venv\Lib\site-packages\numba\np\arraymath.py (3543)


                 File ".venv\Lib\site-packages\numba\np\arraymath.py", line 3543:
                     def diff_impl(a, n=1):
                         <source elided>
                         # To make things easier, normalize input and output into 2d arrays
                         a2 = a.reshape((-1, size))
                         ^

            raised from c:\Users\julvd\repos\transport\.venv\Lib\site-packages\numba\core\typeinfer.py:1091

          During: resolving callee type: Function(<function diff at 0x000001D97FEC5A80>)
          During: typing of call at c:\Users\julvd\repos\transport\.venv\Lib\site-packages\numba\np\arraymath.py (2210)


          File ".venv\Lib\site-packages\numba\np\arraymath.py", line 2210:
                  def impl(x, dx):
                      return np.diff(np.asarray(x))
                      ^

     raised from c:\Users\julvd\repos\transport\.venv\Lib\site-packages\numba\core\typeinfer.py:1091

   During: resolving callee type: Function(<function _get_d at 0x000001D924360400>)
   During: typing of call at c:\Users\julvd\repos\transport\.venv\Lib\site-packages\numba\np\arraymath.py (2227)
   

   File ".venv\Lib\site-packages\numba\np\arraymath.py", line 2227:
       def impl(y, x=None, dx=1.0):
           <source elided>
           yarr = np.asarray(y)
           d = _get_d(x, dx)
           ^

  raised from c:\Users\julvd\repos\transport\.venv\Lib\site-packages\numba\core\typeinfer.py:1091

During: resolving callee type: Function(<function trapz at 0x000001D97FECC720>)
During: typing of call at c:\Users\julvd\repos\transport\mwe.py (11)


File "mwe.py", line 11:
    def __init__(self, aa, xx):
        <source elided>
        self.xx = xx
        self.res = np.trapz(aa, x=self.xx)
        ^

During: resolving callee type: jitclass.Foo#1d904880ad0<xx:array(float64, 1d, A),res:float64>
During: typing of call at <string> (3)

During: resolving callee type: jitclass.Foo#1d904880ad0<xx:array(float64, 1d, A),res:float64>
During: typing of call at <string> (3)


File "<string>", line 3:
<source missing, REPL/exec in use?>

Using xx instead of self.xx works:

import numpy as np
from numba import float64
from numba.experimental import jitclass
@jitclass([
    ("xx" , float64[:]),
    ("res", float64   ),
])
class Foo:
    def __init__(self, aa, xx):
        self.xx = xx
        self.res = np.trapz(aa, x=xx)
aa = np.asarray([1,4,9,16,25], dtype=np.float_)
xx = np.asarray([1,2,3,4,5], dtype=np.float_)
print(Foo(aa, xx).res)

This prints the expected 42.0.

[JulVandenBroeck]

This behavior is not observed when the first argument is an attribute.

The error can be circumvented by creating a copy of the attribute, using np.copy, before using it as the second argument.

[gmarkall]

Thanks for the report - I can reproduce the issue in the original post.

I also note that calling np.broadcast_to with a similar signature even without a jitclass results in a gigantic unintelligible error from type inference:

import numpy as np
from numba import njit, int32, float64


@njit((float64[:], int32))
def Foo(a, n):
    return np.broadcast_to([a], n)

Foo(1., 6)

The last part of the traceback is:

Traceback end

^^^^^
                                              File "/home/gmarkall/numbadev/numba/numba/core/typed_passes.py", line 93, in type_inference_stage
                                                errs = infer.propagate(raise_errors=raise_errors)
                                                       ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                                              File "/home/gmarkall/numbadev/numba/numba/core/typeinfer.py", line 1091, in propagate
                                                raise errors[0]
                                            numba.core.errors.TypingError: Failed in nopython mode pipeline (step: nopython frontend)
                                            No implementation of function Function(<intrinsic np_array>) found for signature:
                                             
                                             >>> np_array(list(array(float64, 1d, A))<iv=None>, none)
                                             
                                            There are 2 candidate implementations:
                                                  - Of which 2 did not match due to:
                                                  Intrinsic in function 'np_array': File: numba/np/arrayobj.py: Line 5396.
                                                    With argument(s): '(list(array(float64, 1d, A))<iv=None>, none)':
                                                   Rejected as the implementation raised a specific error:
                                                     TypingError: array(float64, 1d, A) not allowed in a homogeneous sequence
                                                       Traceback (most recent call last):
                                                         File "/home/gmarkall/numbadev/numba/numba/core/types/functions.py", line 308, in get_call_type
                                                           sig = temp.apply(nolitargs, nolitkws)
                                                                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                                                         File "/home/gmarkall/numbadev/numba/numba/core/typing/templates.py", line 351, in apply
                                                           sig = generic(args, kws)
                                                                 ^^^^^^^^^^^^^^^^^^
                                                         File "/home/gmarkall/numbadev/numba/numba/core/typing/templates.py", line 965, in generic
                                                           result = self._definition_func(self.context, *args, **kws)
                                                                    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                                                         File "/home/gmarkall/numbadev/numba/numba/np/arrayobj.py", line 5399, in np_array
                                                           ret = np_array_typer(typingctx, obj, dtype)
                                                                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                                                         File "/home/gmarkall/numbadev/numba/numba/np/arrayobj.py", line 5386, in np_array_typer
                                                           ndim, seq_dtype = _parse_nested_sequence(typingctx, object)
                                                                             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                                                         File "/home/gmarkall/numbadev/numba/numba/core/typing/npydecl.py", line 484, in _parse_nested_sequence
                                                           n, dtype = _parse_nested_sequence(context, typ.dtype)
                                                                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                                                         File "/home/gmarkall/numbadev/numba/numba/core/typing/npydecl.py", line 482, in _parse_nested_sequence
                                                           raise TypingError("%s not allowed in a homogeneous sequence" % typ)
                                                       numba.core.errors.TypingError: array(float64, 1d, A) not allowed in a homogeneous sequence
                                              raised from /home/gmarkall/numbadev/numba/numba/core/typing/npydecl.py:482
                                            
                                            During: resolving callee type: Function(<intrinsic np_array>)
                                            During: typing of call at /home/gmarkall/numbadev/numba/numba/np/arrayobj.py (5433)
                                            
                                            
                                            File "../../numba/numba/np/arrayobj.py", line 5433:
                                                def impl(object, dtype=None):
                                                    return np_array(object, dtype)
                                                    ^
                                   raised from /home/gmarkall/numbadev/numba/numba/core/typeinfer.py:1091
                                 
                                 During: resolving callee type: Function(<built-in function array>)
                                 During: typing of call at /home/gmarkall/numbadev/numba/numba/np/arraymath.py (4312)
                                 
                                 
                                 File "../../numba/numba/np/arraymath.py", line 4312:
                                             def impl(a, dtype=None):
                                                 return np.array(a)
                                                 ^
                        raised from /home/gmarkall/numbadev/numba/numba/core/typeinfer.py:1091
                      
                      During: resolving callee type: Function(<built-in function asarray>)
                      During: typing of call at /home/gmarkall/numbadev/numba/numba/np/arrayobj.py (1457)
                      
                      
                      File "../../numba/numba/np/arrayobj.py", line 1457:
                      def _default_broadcast_to_impl(array, shape):
                          array = np.asarray(array)
                          ^
             raised from /home/gmarkall/numbadev/numba/numba/core/typeinfer.py:1091
           
           During: resolving callee type: Function(<function broadcast_to at 0x77079a6bdf80>)
           During: typing of call at /home/gmarkall/numbadev/numba/numba/np/arrayobj.py (1470)
           
           
           File "../../numba/numba/np/arrayobj.py", line 1470:
                   def impl(array, shape):
                       return np.broadcast_to(array, (shape,))
                       ^
  raised from /home/gmarkall/numbadev/numba/numba/core/typeinfer.py:1091

During: resolving callee type: Function(<function broadcast_to at 0x77079a6bdf80>)
During: typing of call at /home/gmarkall/numbadev/issues/9522/repro.py (7)


File "repro.py", line 7:
def Foo(a, n):
    return np.broadcast_to([a], n)
    ^

[gmarkall]

Labelling as a bug for now, but I think a bit of discussion in the triage meeting will help establish the right directions to take for looking further into this bug / this collection of related issues.

[gmarkall]

After discussion in the triage meeting - the issue is that broadcasting a list is not supported - if you replace the list that's the first argument of broadcast_to with a tuple, does that work for your use case?

[guilhermeleobas]

@JulVandenBroeck, are you trying to broadcast a set of arrays? If so, can you try np.broadcast_arrays, instead? np.broadcast_to expects the first argument to be an array-like, that's why you're seeing the error.

[JulVandenBroeck]

After discussion in the triage meeting - the issue is that broadcasting a list is not supported - if you replace the list that's the first argument of broadcast_to with a tuple, does that work for your use case?

Thanks for the info! Using a tuple gives the same error, and so does an ndarray. It finds the appropriate candidate implementations but fails during the native lowering step.

@JulVandenBroeck, are you trying to broadcast a set of arrays? If so, can you try np.broadcast_arrays, instead? np.broadcast_to expects the first argument to be an array-like, that's why you're seeing the error.

No, I want to broadcast a single array to a certain shape: in my example an array with shape (1,) to shape (5,). I used [a] as the first argument, which is an array-like. From the traceback, I deduce that Numba finds the correct implementation but fails to compile. I personally believe it has nothing to do with Numpy.

[JulVandenBroeck]

I decided to switch to a structref, which solved all my issues (for the moment being :) )