Efficiency for calling Julia from python and purely run Julia

[Geositta]

Thank you so much for pointing out.

The motivation is, from I heard and some tests,

Is there any way to optimize array additions and multiplications with transposes?

I checked the following Python script import numpy as np import time n0=n1=n2=n3=30 A = np.random.random((n0,n1,n2,n3)) n_loop = 100 start = time.time() for i in range(n_loop): B = np.add( 0.1*A, 0.2*np.transpose(A,(1,0,2,3)) ) end = time.time() print((end - start)/n_loop * 1000, " ms") against Julia’s using BenchmarkTools, LoopVectorization, Tullio function test1(A) B = similar(A) for i in 1:size(A,1), j in 1:size(A,2) # B[i,j,:,:] = 0.1*A[i,j,:,:] + 0.2*A[i,j,:,:]' …

TensorOperations.jl is often the fastest way to do permutedims on larger arrays. It has a smarter cache-friendly blocking algorithm than the one in Base. But how much this matters of course depends on size & permutation.

My goal involves a series of array additions and permutations, it is complicated to optimize memory cache to improve the performance in Numpy/Fortran. TensorOperations.jl has some feature on it.

In one example in the above link, the timing from python is 4.956309795379639 ms and in julia is 2.585 ms (21 allocations: 6.18 MiB), about a factor of 2 better. I observed a similar pattern in other cases.