Bug with af.select when using the OpenCL backend

[Shyam Sundar Sankaran]

There seems to be a bug with the af.select() function operating on spliced 2D arrays, when using the OpenCL backend. I've illustrated this with the code below:

af.set_backend("opencl")

x = af.to_array(np.arange(-5,6))
x = af.tile(x, 1, 2)

x[2:-2] = af.select(x[2:-2]>=0, x[2:-2], -1*x[2:-2])

print(x)

This gives the output as:

arrayfire.Array()
Type: long int

[11 2 1 1]
        -5         -5 
        -4         -4 
         3          3 
         2          2 
         1          1 
         0         -4 
         1         -3 
         2         -2 
         3         -1 
         4          4 
         5          5 

Changing the backend to CPU, however gives the expected answer:

arrayfire.Array()
Type: long int

[11 2 1 1]
        -5         -5 
        -4         -4 
         3          3 
         2          2 
         1          1 
         0          0 
         1          1 
         2          2 
         3          3 
         4          4 
         5          5 

However if I copy the spliced values into a new array, the output generated by the OpenCL backend is also correct. That is:

af.set_backend("opencl")
x = af.to_array(np.arange(-5,6))
x = af.tile(x, 1, 2)

x_temp  = x[2:-2].copy()
x_temp  = af.select(x_temp>=0, x_temp, -1*x_temp)
x[2:-2] = x_temp

print(x)

[Pavan Yalamanchili]

@ShyamSS-95 for now can you use the following as a work around

cond = (x[2:-2] >= 0).as(af.Dtype.s32)
x[2:-2] =  (2 * cond - 1) * x[2:-2] #cond * x[2:-2] + (cond - 1) * x[2:-2]

[Pavan Yalamanchili]

btw this is a bug in arrayfire. I have reported it over here: arrayfire/arrayfire#1730

[Pavan Yalamanchili]

@ShyamSS-95 I fixed it in this PR: arrayfire/arrayfire#1731

You can pull those changes and rebuild arrayfire if you don't want to wait.

[Pavan Yalamanchili]

@ShyamSS-95 This is fixed in upstream arrayfire.