arrayfire seems a bit slower than cupy

[xichaoqiang]

I compare the compute speed with arrayfire,cupy,and numpy.

The resuts show cupy is a bit faster than arrayfire.

Is there another way to conver af.array to np.array?

Is the speed of convert of af.array to np.array slow?

import arrayfire as af
import numpy as np
import cupy as cp
from array import array
from time import time
import numpy.random_intel as rndi

N=2000
ru=rndi.random_sample((N,N)).astype(np.float32)
x_af=af.np_to_af_array(ru)
x_np=ru.copy()
x_cp=cp.array(ru)

print("N=",N)
t0=time();r1=af.matmul(x_af,x_af);r1=np.array(r1);t1=time()-t0
print("af t1=",t1)
t0=time();r2=af.matmul(x_af+1.0,x_af+1.0);r2=np.array(r2);t2=time()-t0
print("af t2=",t2)
t0=time();r3=cp.matmul(x_cp,x_cp);r3=cp.asnumpy(r3);t3=time()-t0
print("cp t3=",t3)
t0=time();r4=cp.matmul(x_cp+1.0,x_cp+1.0);r4=cp.asnumpy(r4);t4=time()-t0
print("cp t4=",t4)
t0=time();r5=np.matmul(x_np,x_np);t5=time()-t0
print("np t5=",t5)
t0=time();r6=np.matmul(x_np+1.0,x_np+1.0);t6=time()-t0
print("np t6=",t6)

Resut shows that.

N= 2000
af t1= 0.030921220779418945
af t2= 0.03490591049194336
cp t3= 0.016954898834228516
cp t4= 0.01795172691345215
np t5= 0.10272407531738281
np t6= 0.11370682716369629

N= 3000
af t1= 0.2603302001953125
af t2= 0.19445013999938965
cp t3= 0.061836957931518555
cp t4= 0.05884242057800293
np t5= 0.27725768089294434
np t6= 0.3281223773956299

N= 5000
af t1= 0.5754146575927734
af t2= 0.3031926155090332
cp t3= 0.20644783973693848
cp t4= 0.19946646690368652
np t5= 1.2706019878387451
np t6= 1.3493928909301758

N= 8000
af t1= 0.9365062713623047
af t2= 0.7700843811035156
cp t3= 0.5515463352203369
cp t4= 0.5555505752563477
np t5= 5.050236940383911
np t6= 5.4347453117370605

N= 10000
af t1= 1.2895522117614746
af t2= 1.3114979267120361
cp t3= 0.9285182952880859
cp t4= 0.9394872188568115
np t5= 9.278713941574097
np t6= 9.743171691894531

[xichaoqiang]

When I try the .to_ndarray() function,the speed of arrarfire is faster than before.