Binary morphology: omit weights array when possible

[grlee77]

This is a follow-up to #485 allowing the overhead of creating a weights array of uniform values to be omitted when possible for all binary morphology operators. This gives a large performance benefit for small footprints (and small images) with the benefit becoming minimal for large footprint size.

As currently implemented, a shape tuple can be passed to structure to indicate the equivalent of an array of ones of that shape. This is a deviation from SciPy's API which does not currently allow this. I am not positive if this is the best API choice or if it might be better to instead add a size keyword-only argument like the grayscale morphology functions currently have.

The footprint generation methods square, rectangle, cube etc. use this new tuple format when possible to reduce overhead.

I will post benchmarks below.

[grlee77]

Benchmark results

Acceleration here is the ratio of the runtime when passing a boolean array of footprint_shape size vs. just passing the size tuple.

These are worse case accelerations as they assume the footprint had already been created and don't include the initial overhead of allocating the footprint on the GPU.

Acceleration factors run from ~3x for small images approaching 1x as the image size becomes larger.

shape	footprint_shape	acceleration
(256, 256)	(3, 3)	3.04
(256, 256)	(5, 5)	3.06
(256, 256)	(7, 7)	2.90
(256, 256)	(15, 15)	2.91
(256, 256)	(31, 31)	2.88
(512, 512)	(3, 3)	3.03
(512, 512)	(5, 5)	2.77
(512, 512)	(7, 7)	2.88
(512, 512)	(15, 15)	2.89
(512, 512)	(31, 31)	2.82
(1024, 1024)	(3, 3)	2.57
(1024, 1024)	(5, 5)	2.22
(1024, 1024)	(7, 7)	2.73
(1024, 1024)	(15, 15)	2.58
(1024, 1024)	(31, 31)	2.08
(2048, 2048)	(3, 3)	1.83
(2048, 2048)	(5, 5)	1.55
(2048, 2048)	(7, 7)	1.66
(2048, 2048)	(15, 15)	1.47
(2048, 2048)	(31, 31)	1.31
(4096, 4096)	(3, 3)	1.25
(4096, 4096)	(5, 5)	1.14
(4096, 4096)	(7, 7)	1.18
(4096, 4096)	(15, 15)	1.12
(4096, 4096)	(31, 31)	1.08
(32, 32, 32)	(3, 3, 3)	2.99
(32, 32, 32)	(5, 5, 5)	2.87
(32, 32, 32)	(9, 9, 9)	2.88
(32, 32, 32)	(13, 13, 13)	2.86
(64, 64, 64)	(3, 3, 3)	2.63
(64, 64, 64)	(5, 5, 5)	2.85
(64, 64, 64)	(9, 9, 9)	2.89
(64, 64, 64)	(13, 13, 13)	2.87
(128, 128, 128)	(3, 3, 3)	1.63
(128, 128, 128)	(5, 5, 5)	2.17
(128, 128, 128)	(9, 9, 9)	1.94
(128, 128, 128)	(13, 13, 13)	1.83
(256, 256, 256)	(3, 3, 3)	1.09
(256, 256, 256)	(5, 5, 5)	1.12
(256, 256, 256)	(9, 9, 9)	1.10
(256, 256, 256)	(13, 13, 13)	1.09

[gigony]

Thanks @grlee77 ! Looks good to me!

[gigony]

unnecessary blank line?

[grlee77]

rerun tests

[grlee77]

@gigony, test failures in unrelated isotropic morphology functions have been fixed. I added checks to enforce properties of the provided output.

[grlee77]

/merge