Conversion from af:Array to scalar

[Shyam Sundar Sankaran]

Is there a Python command to convert an ArrayFire array to a scalar?

Suppose I define a = af.Array([1,2,3]). Then, a[1] gives me an af.Array of dimensions [1 1 1 1] containing the value of 2. But how do I get the scalar value of it directly? I'm currently using af.algorithm.sum(a[1])

Is there a more direct way to do this?

[Pavan Yalamanchili]

@ShyamSS-95 I just realized that is the fastest way for now. The C++ API has array.scalar<T>() but that is copying the entire array to the host and then returning one element.

Created an issue upstream to add the necessry functionality: arrayfire/arrayfire#1725

[Pavan Yalamanchili]

@ShyamSS-95 Actually in your case, since a[1] is just one element, you could do a[1].to_array() to get a host array of size [1] or a[1].to_array()[0] to get a number.

Can you also tell in what scenario you are using this ? Because if you are copying to host costantly, it is better off if you do the copy once and then index the host instead of copying each element individually.

[Shyam Sundar Sankaran]

I'm using it in a loop for the implementation of a collision algorithm. Shown below is the code snippet where I've had to convert an af.Array to a scalar.

for i in range(no_of_particles):
  x_particle_considered = af.data.constant(af.algorithm.sum(x_coordinates[i]), (no_of_particles-1-i))
  y_particle_considered = af.data.constant(af.algorithm.sum(y_coordinates[i]), (no_of_particles-1-i))  
  x_coordinates_others  = x_coordinates[(i+1):no_of_particles]        
  y_coordinates_others  = y_coordinates[(i+1):no_of_particles]
  # Determining the components of the vector joining particle i to all other particles    
  x_coordinates_others = x_coordinates_others-x_particle_considered
  y_coordinates_others = y_coordinates_others-y_particle_considered

[Pavan Yalamanchili]

@ShyamSS-95 try this:

@af.broadcast
def sub(lhs, rhs):
    return lhs - rhhs

for i in range(no_of_particles):
  x_coordinates_others  = x_coordinates[(i+1):no_of_particles]        
  y_coordinates_others  = y_coordinates[(i+1):no_of_particles]
  # Determining the components of the vector joining particle i to all other particles    
  x_coordinates_others = sub(x_coordinates_others, x_coordinates[i])
  y_coordinates_others = sub(y_coordinates_others, y_coordinates[i])

[Shyam Sundar Sankaran]

Thanks! It's much easier to write. I wrote the following code to check the effect on speed:

import numpy as np 
import arrayfire as af
import time

no_of_particles = 1000
test_array = np.random.rand(no_of_particles)
x_coordinates = y_coordinates = af.Array(test_array.ctypes.data, test_array.shape, test_array.dtype.char)

@af.broadcast
def sub(lhs, rhs):
  return lhs - rhs

sum_time_start = time.time()
for i in range(no_of_particles-1):
  x_coordinates_others = x_coordinates[(i+1):no_of_particles]        
  y_coordinates_others = y_coordinates[(i+1):no_of_particles]
  x_coordinates_others = x_coordinates_others - af.data.constant(af.algorithm.sum(x_coordinates[i]), (no_of_particles-1-i))
  y_coordinates_others = y_coordinates_others - af.data.constant(af.algorithm.sum(y_coordinates[i]), (no_of_particles-1-i))
sum_time_end  = time.time()

print("Time elapsed using sum = ", (sum_time_end- sum_time_start) )

sub_time_start = time.time()
for i in range(no_of_particles-1):
  x_coordinates_others = x_coordinates[(i+1):no_of_particles]        
  y_coordinates_others = y_coordinates[(i+1):no_of_particles]
  x_coordinates_others = sub(x_coordinates_others, x_coordinates[i])
  y_coordinates_others = sub(y_coordinates_others, y_coordinates[i])
sub_time_end  = time.time()

print("Time elapsed using sub = ", (sub_time_end- sub_time_start) )

It runs faster as well:

Time elapsed using sum =  0.628760576248169
Time elapsed using sub  =  0.25028181076049805

[Pavan Yalamanchili]

@ShyamSS-95 excellent 👍

af.broadcast allows you to perform elemnt wise operations where the dimensions are not identical. But when you are benchmarking, you need to have af.eval(x_coordinates_others, y_coordinates_others) at the end of the loop and an af.sync() after the loop.