1

I'm trying to learn the Arrayfire idioms by translating some vectorised numpy code.

For example, this is valid rowwise addition and multiplication in numpy,

>>> a = np.array([1,2,3])
>>> a
array([1, 2, 3])
>>> b = np.arange(9).reshape((3, 3))
>>> b
array([[0, 1, 2],
       [3, 4, 5],
       [6, 7, 8]])
>>> a + b
array([[ 1,  3,  5],
       [ 4,  6,  8],
       [ 7,  9, 11]])
>>> a * b
array([[ 0,  2,  6],
       [ 3,  8, 15],
       [ 6, 14, 24]])

Do all arrays in arrayfire need to be the same shape? This generates an error.

>>> a = af.from_ndarray(a)
>>> b = af.from_ndarray(b)
>>> a + b
Invalid dimension for argument 1
Expected: ldims == rdims
>>> a * b
Invalid dimension for argument 1
Expected: ldims == rdims

From @pradeep answer, you can do this using tile until the added as a feature request.

>>> a = np.array([1,2,3])
>>> a = af.tile(af.transpose(af.from_ndarray(a)),3,1)
>>> af.display(a)

[3 3 1 1]
         1          2          3 
         1          2          3 
         1          2          3 

>>> b = np.arange(9).reshape((3, 3))
>>> b = af.from_ndarray(b)
>>> af.display(a + b)

[3 3 1 1]
         1          3          5 
         4          6          8 
         7          9         11 

>>> af.display(a * b)

[3 3 1 1]
         0          2          6 
         3          8         15 
         6         14         24
Daniel Farrell
  • 9,316
  • 8
  • 39
  • 62

1 Answers1

2

As of writing this response, yes that is correct. Array's need to be of same shape. But I would like to point out that we are already working on broadcasting feature for binary operations - here is the PR - we will try to get this feature into a release as soon as we can.

However, even with current release, this limitation can be easily worked around using tile function. Since tile will be a JIT operation for such broadcast operations, it won't allocate any additional memory. The arithmetic operation and tiling operation will be combined into an efficient single launch kernel.

pradeep
  • 712
  • 4
  • 13