Refactored Perceptionhash for performance

[evanoberholster]

Refactored PerceptionHash for performance. If you want me to change the name to replace existing function that would be acceptable.

Features:

• Significantly reduces allocations
• Uses static DCT1D tables
• Uses sync pool to reduce pixel data allocations
• Uses image.Image optimizations for obtaining pixels for improved inlining of functions
• Backward compatible with PerceptionHash, they are interchangeable. Confirmed with passing tests.
• Performance without image resize as below.

name                      time/op
PerceptionHash/Regular-12         358µs ±26%
PerceptionHash/Fast-12           84.8µs ± 7%
PerceptionHash/Parallel-12       82.5µs ± 3%
PerceptionHash/Fast-Parallel-12  11.5µs ± 1%

name                      alloc/op
PerceptionHash/Regular-12         193kB ± 0%
PerceptionHash/Fast-12            17.0B ± 0%
PerceptionHash/Parallel-12        193kB ± 0%
PerceptionHash/Fast-Parallel-12    198B ± 0%

name                      allocs/op
PerceptionHash/Regular-12         4.68k ± 0%
PerceptionHash/Fast-12             0.00     
PerceptionHash/Parallel-12        4.68k ± 0%
PerceptionHash/Fast-Parallel-12    0.00 

[evanoberholster]

Comparison between PerceptionHash and PerceptionHashFast:

name               old time/op    new time/op    delta
PerceptionHash-12     945µs ± 9%     689µs ± 2%  -27.08%  (p=0.000 n=170+177)

name               old alloc/op   new alloc/op   delta
PerceptionHash-12     471kB ± 0%     298kB ± 0%  -36.80%  (p=0.000 n=174+198)

name               old allocs/op  new allocs/op  delta
PerceptionHash-12     4.74k ± 0%     0.07k ± 0%  -98.61%  (p=0.000 n=200+200)

[corona10]

@evanoberholster Looks nice! Thank you for your contribution!!

If you want me to change the name to replace existing function that would be acceptable.

I prefer this one(change the name ) :)

[corona10]

Please add your name to the following file.
https://github.com/corona10/goimagehash/blob/master/AUTHORS.md

[corona10]

Significantly reduces allocations
Uses static DCT1D tables
Uses sync pool to reduce pixel data allocations
Uses image.Image optimizations for obtaining pixels for improved inlining of functions
Backward compatible with PerceptionHash, they are interchangeable. Confirmed with passing tests.
Performance without image resize as below.

Thanks, I love all of your changes :)

[corona10]

Would you like to switch the PerceptionHashFast into PerceptionHash and remove old implementation(including old dct implementation)?

[corona10]

Suggested change

	func PerceptionHashFast(img image.Image) (*ImageHash, error) {
	func PerceptionHash(img image.Image) (*ImageHash, error) {

[evanoberholster]

Made these changes in new update.

[evanoberholster]

PerceptionHashFast DCT static tables only support a size of 64x64. We would need to include further static tables to support other image sizes, (eg 256x256)

[corona10]

@evanoberholster

PerceptionHashFast DCT static tables only support a size of 64x64.

Ah okay, let's replace PerceptionHash only.

[evanoberholster]

Sorry for the delay I have been busy at work. Will get it replaced and update the PR.

[corona10]

Sorry for the delay I have been busy at work. Will get it replaced and update the PR.

Yeah, Don't worry I can wait :)

[corona10]

Almost looks good to me, Just 1 nit comment and I will take a look at more later :)
Thanks for the change!

[evanoberholster]

I agree that those tests are redundant. Thanks.