perf: IBA::resample improvements to speed up 20x or more

[lgritz]

For IBA::resample() when bilinear interpolation is used, almost all of the expense was due to its relying on ImageBuf::interppixel which is simple but constructs a new ImageBuf::ConstIterator EVERY TIME, which is very expensive.

Reimplement in a way that reuses a single iterator. This speeds up IBA::resample by 20x or more typicaly.

Also refactor resample to pull the handling of deep images into a separate helper function and out of the main inner loop. And add some benchmarking.

[ssh4net]

I have added your fixes to hwy branch and run a comparison. It now looks more fair :)

[ Resample 75% ]
uint8      |       9.43 |       1.67 |   5.65x
uint16     |       9.17 |       1.81 |   5.07x
uint32     |      26.92 |      19.57 |   1.38x
float      |       9.55 |       1.87 |   5.11x
half       |      11.30 |       4.69 |   2.41x
double     |      32.98 |      30.87 |   1.07x

[ Resample 50% ]
uint8      |       4.20 |       0.79 |   5.31x
uint16     |       4.19 |       0.89 |   4.70x
uint32     |      17.40 |      17.01 |   1.02x
float      |       4.33 |       0.79 |   5.48x
half       |       5.36 |       2.25 |   2.38x
double     |      20.27 |      18.89 |   1.07x

[ Resample 25% ]
uint8      |       1.24 |       0.24 |   5.05x
uint16     |       1.19 |       0.26 |   4.52x
uint32     |      11.98 |      11.16 |   1.07x
float      |       1.08 |       0.20 |   5.39x
half       |       1.29 |       0.55 |   2.37x
double     |      16.50 |      13.57 |   1.22x
Type       | Scalar(ms) |   SIMD(ms) | Speedup

[jessey-git]

Ran some tests locally and can also confirm the perf goodness (on boring uint8 pngs) as well as successful idiff comparisons of main vs the PR.

Time in ms average over 5 attempts:

main	PR
4096x4096, 25% --> 1024x1024 : Avg 282.5024	4096x4096, 25% --> 1024x1024 : Avg 16.9484
4096x4096, 33% --> 1351x1351 : Avg 484.7668	4096x4096, 33% --> 1351x1351 : Avg 29.4051
4096x4096, 50% --> 2048x2048 : Avg 1126.7262	4096x4096, 50% --> 2048x2048 : Avg 61.9360
4096x4096, 67% --> 2744x2744 : Avg 2021.3136	4096x4096, 67% --> 2744x2744 : Avg 96.1518
4096x4096, 75% --> 3072x3072 : Avg 2529.3684	4096x4096, 75% --> 3072x3072 : Avg 124.1163
3840x2160, 25% --> 960x540 : Avg 139.5478	3840x2160, 25% --> 960x540 : Avg 7.0885
3840x2160, 33% --> 1267x712 : Avg 243.2302	3840x2160, 33% --> 1267x712 : Avg 12.2748
3840x2160, 50% --> 1920x1080 : Avg 558.0922	3840x2160, 50% --> 1920x1080 : Avg 25.5491
3840x2160, 67% --> 2572x1447 : Avg 996.6915	3840x2160, 67% --> 2572x1447 : Avg 47.6503
3840x2160, 75% --> 2880x1620 : Avg 1247.9340	3840x2160, 75% --> 2880x1620 : Avg 62.8001
2160x3840, 25% --> 540x960 : Avg 139.3882	2160x3840, 25% --> 540x960 : Avg 8.7678
2160x3840, 33% --> 712x1267 : Avg 244.5914	2160x3840, 33% --> 712x1267 : Avg 16.3722
2160x3840, 50% --> 1080x1920 : Avg 557.2218	2160x3840, 50% --> 1080x1920 : Avg 25.6951
2160x3840, 67% --> 1447x2572 : Avg 994.1367	2160x3840, 67% --> 1447x2572 : Avg 45.7704
2160x3840, 75% --> 1620x2880 : Avg 1246.6437	2160x3840, 75% --> 1620x2880 : Avg 60.2098
116x2052, 25% --> 29x513 : Avg 4.1224	116x2052, 25% --> 29x513 : Avg 0.3348
116x2052, 33% --> 38x677 : Avg 7.1067	116x2052, 33% --> 38x677 : Avg 0.4689
116x2052, 50% --> 58x1026 : Avg 16.1875	116x2052, 50% --> 58x1026 : Avg 0.8663
116x2052, 67% --> 77x1374 : Avg 28.1018	116x2052, 67% --> 77x1374 : Avg 1.4981
116x2052, 75% --> 87x1539 : Avg 36.0132	116x2052, 75% --> 87x1539 : Avg 2.4759

[lgritz]

I have added your fixes to hwy branch and run a comparison. It now looks more fair :)

Yeah, around 5x speedup is what I'd expect, or at least hope for, for a good scalar vs good AVX2 implementation. Very far from that in either direction tells me that one of them has something wrong with it that is probably easily fixable.

[lgritz]

Approval on this?

[ssh4net]

If no regression in results, then for sure it ready :)