Poor slicing performance compared to NumPy

Beginnen Sie damit, den mitgelieferten Python-Benchmark mit den CPU- und GPU-Backends auszuführen, wobei Sie sich auf af_B[:, i], af_B[i, :] und af.matmul konzentrieren. Da in der Issue keine Quelldatei und kein Test genannt werden, verfolgen Sie diese Einstiegspunkte durch die Python-Bindings und vergleichen Sie sie mit unsliced matmul. Als abgeschlossen gilt die Aufgabe, wenn Sie die Regression reproduziert, ihre Ursache identifiziert und einen Regressionstest oder Benchmark hinzugefügt haben, der die Verbesserung demonstriert.

reported by @floopcz on over here: https://github.com/arrayfire/arrayfire/issues/1428

ArrayFire slicing seems to suffer from a performance issue. Consider the following python code, that:

• calculates the dot product of two matrices, first using NumPy, than ArrayFire
• calculates each column/row of the dot product separately by slicing a single column/row from one of the matrices

#!/usr/bin/env python3
from time import time
import arrayfire as af
import numpy as np

af.set_backend('cpu')
af.info()

iters = 1000
n = 512

af_A = af.randu(n, n)
af_B = af.randu(n, n)

np_A = np.random.rand(n, n).astype(np.float32)
np_B = np.random.rand(n, n).astype(np.float32)

start = time()
for t in range(iters):
    np_C = np.dot(np_A, np_B)
print('numpy - dot: {}'.format(time() - start))

af.sync()
start = time()
for t in range(iters):
    af_C = af.matmul(af_A, af_B)
af.sync()
print('arrayfire - matmul: {}'.format(time() - start))

start = time()
for t in range(iters):
    for i in range(np_B.shape[1]):
        np_C = np.dot(np_A, np_B[:, i])
print('numpy - sliced dot - column major: {}'.format(time() - start))

af.sync()
start = time()
for t in range(iters):
    for i in range(af_B.shape[1]):
        af_C = af.matmul(af_A, af_B[:, i])
af.sync()
print('arrayfire - sliced matmul - column major: {}'.format(time() - start))

start = time()
for t in range(iters):
    for i in range(np_B.shape[0]):
        np_C = np.dot(np_B[i, :], np_A)
print('numpy - sliced dot - row major: {}'.format(time() - start))

af.sync()
start = time()
for t in range(iters):
    for i in range(af_B.shape[0]):
        af_C = af.matmul(af_B[i, :], af_A)
af.sync()
print('arrayfire - sliced matmul - row major: {}'.format(time() - start))

The results are following:

ArrayFire v3.3.2 (CPU, 64-bit Linux, build f65dd97)
[0] Unknown: Unknown, 15880 MB, Max threads(1) 
numpy - dot: 1.3848536014556885
arrayfire - matmul: 1.325775146484375
numpy - sliced dot - column major: 7.156768798828125
arrayfire - sliced matmul - column major: 38.87605834007263
numpy - sliced dot - row major: 7.6784679889678955
arrayfire - sliced matmul - row major: 41.27544379234314

The results suggest that with slicing, arrayfire performance is significantly degraded compared to NumPy. I have achieved similarly distributed results also with the GPU backend. Both numpy and arrayfire are linked against Intel MKL.

Am I doing something "illegal" or is it an inefficiency of the library? Thanks.