Problem with af.mean in parallel loop

[Matus Krajnak]

When I run

data3d = af.randu(1,1,5)
means  = af.constant(0,5)
meansP = af.constant(0,5)

for n in range(5):
    means[n]=af.mean(data3d[:,:,n])
    
for n in af.ParallelRange(5):
    meansP[n]=af.mean(data3d[:,:,n])
    
af.sync

print(means)
print(meansP)

results in:

arrayfire.Array()
Type: float

[5 1 1 1]
0.4703
0.5648
0.2519
0.2058
0.3630

arrayfire.Array()
Type: float

[5 1 1 1]
0.3712
0.3712
0.3712
0.3712
0.3712

[Pavan Yalamanchili]

You can't use a function that returns a scalar inside parallel for.

You can achieve the same thing by doing means = af.mean(data, dim=2)

[Matus Krajnak]

thank you @pavanky, also is this error expected?

diff=0.0
for x in range(10000):
    data3d = af.randn(100,100,2)
    data2d = af.moddims(data3d,10000,2)
    diff+=np.abs(
        1.0-af.mean(data2d[:,0])/af.mean(data2d,dim=0)[0,0])
print('Difference = ',diff/10000)

Difference = [1.4164448e-07]

[Pavan Yalamanchili]

@matkraj yes the number of threads launched for af.mean(data2d[:, 0]) is different from af.mean(data2d, dim=0). This results in small floating point differences because floating point arithmetic is not associative.

If you used something like af.min or af.max which dont perform floating point arithmetic you will see identical results.

[Matus Krajnak]

Thanks!