Implement SimpleArray member functions for searching part 2

[pchsu-hsupc]

Completed Tasks

• Make Python unittest for SimpleArray::argwhere() and SimpleArray::where()

• Using mixin class SimpleArrayMixinConditional to collect methods SimpleArray::argwhere() and SimpleArray::where()

Abstraction

As mentioned in #519 , this PR implements argwhere() and where() methods without overloading operators.

numpy.argwhere(a): Find the indices of array elements that are non-zero by default, also could add condition as well.

numpy.where(condition, [x, y, ]/): Return elements chosen from x or y depending on condition.

Implement Method

Thanks to the std::function template, I could use lambdas as argument—letting the API mirror NumPy’s style in Python.

ret_np = np.argwhere(narr == 5)
ret_sa = sarr.argwhere(lambda x: x == 5)

[pchsu-hsupc]

• Use an std::vector to store the indices of elements that satisfy the condition.
• Creating the output coordinates and filling the dimensions's information.

[pchsu-hsupc]

• Should output the same shape as input.
• If the condition is met, fill with the original value; otherwise, use the alternative.

[pchsu-hsupc]

Numpy's where() says that when only condition is provided, nonzero() should be preferred. So I didn't give SimpleArray::where default operation.

[pchsu-hsupc]

Thank you, @ThreeMonth03 and @KHLee529, for your valuable insights.
I would appreciate any additional feedback you might have.
@yungyuc, could you please take a look at this PR when you have a moment? Thanks!

[yungyuc]

Could you please profile the runtime performance of the searching functions like what issue #532 does? If the runtime is much worse than numpy, we may not use the implementation in this PR and need to find another approach.

[pchsu-hsupc]

Sure, I will add it in.

[pchsu-hsupc]

Profiling

Use the built-in tools modmesh.call_profiler and modmesh.time_registry to profile the default behavior of SimpleArray::argwhere() from Python. This function is equivalent to SimpleArray::argwhere(lambda x: x != 0).

Numpy is abbreviated to np ,SimpleArray is abbreviated to sa.
And Nums is the abbreviation of number of elements.
Speedup achieved by SimpleArray is measured relative to NumPy.

	np.argwhere() per call	sa.argwhere() per call	Speedup
Nums = 128, Data type = uint8	5.489 ms	1.849 ms	2.97x
Nums = 1K, Data type = uint16	3.150 ms	3.795 ms	0.83x
Nums = 8K, Data type = uint16	11.546 ms	24.759 ms	0.47x
Nums = 64K, Data type = uint32	80.181 ms	198.060 ms	0.40x
Nums = 512K, Data type = uint32	980.167 ms	2165.842 ms	0.45x
Nums = 4M, Data type = uint64	9604.355 ms	17401.239 ms	0.55x
Nums = 32M, Data type = uint64	102880.127 ms	160952.633 ms	0.64x

With iterations = 50, the test cases were generated using np.random.uniform().

[yungyuc]

Please include your profile script in the PR. @KHLee529 could you please help @pchsu-hsupc ?

The runtime is much less slower than I thought, but still too slow to be useful. We may not implement the interface by using a Python callable and/or lambda.

[KHLee529]

@pchsu-hsupc I think it 's worthy to first clarify whether it's the Python callable value that makes runtime slow or the C++ std::function. Maybe try to make the default argument condition a free function written in C++. If it is the Python callable object that slows down the whole process, we should implement this function in another way.

[yungyuc]

@pchsu-hsupc In the discussions in the meetup, we concluded that it should use Boolean array to select elements in the way that numpy employs. Please update your progress.