Use an unordered_set in thin_with_voxel

Jean-Romain commented 5 years ago

@tiagodc could you benchmark that on TLS data.

tiagodc commented 5 years ago

sure

tiagodc commented 5 years ago

`-thin_with_voxel` TLS benchmark using `set`

CPU: Intel(R) Core(TM) i7-8550U CPU @ 1.80GHz
cores: 8 (4 physical x 4 threads)
RAM: 16 GB
average CPU usage during file streaming: 10%
reading function: lidR::readLAS

Cloud 1

N: 2,541,917 points
extents: 30 x 30 x 14 m

Voxel spacing (m)	surviving points	read time (s)	~ memory usage (GB)
0.1	349,787	4	0.06
0.05	925,881	5	0.12
0.025	1,718,986	7	0.18
0.01	2,414,604	7	0.25

Cloud 2

N: 10,773,659 points
extents: 86 x 127 x 19 m

Voxel spacing (m)	surviving points	read time (s)	~ memory usage (GB)
0.1	1,049,797	16	0.11
0.05	2,907,514	18	0.23
0.025	6,252,097	28	0.51
0.01	9,947,220	32	0.99

Cloud 3

N: 42,869,489 points
extents: 400 x 217 x 56 m

Voxel spacing (m)	surviving points	read time (s)	~ memory usage (GB)
0.1	2,606,613	91	0.21
0.05	7,602,382	104	0.67
0.025	19,556,071	126	1.62
0.01	37,853,465	138	3.18

Cloud 4

N: 165,155,435 points
extents: 259 x 234 x 193 m

Voxel spacing (m)	surviving points	read time (s)	~ memory usage (GB)
0.1	1,506,216	187	0.31
0.05	4,552,934	208	0.87
0.025	12,987,644	231	1.14
0.01	41,461,462	309	3.50

Jean-Romain commented 5 years ago

Thank you Tiago but what I was expecting was a comparison between set and unordered_set. On ALS data I benchmarked a gain of ~33%

tiagodc commented 5 years ago

oh crap... I didn't see you created a new branch for that. OK, I'll pull the unordered_set branch, recompile and perform the tests again... no problem

tiagodc commented 5 years ago

`-thin_with_voxel` TLS benchmark using `unordered_set`

CPU: Intel(R) Core(TM) i7-8550U CPU @ 1.80GHz
cores: 8 (4 physical x 4 threads)
RAM: 16 GB
average CPU usage during file streaming: 25% - last time my PC was under power management
reading function: lidR::readLAS

Cloud 1

N: 2,541,917 points
extents: 30 x 30 x 14 m

Voxel spacing (m)	surviving points	read time (s)	~ memory usage (GB)
0.1	349,787	1	0.03
0.05	925,881	1	0.12
0.025	1,718,986	2	0.14
0.01	2,414,604	2	0.23

Cloud 2

N: 10,773,659 points
extents: 86 x 127 x 19 m

Voxel spacing (m)	surviving points	read time (s)	~ memory usage (GB)
0.1	1,049,797	6	0.28
0.05	2,907,514	7	0.38
0.025	6,252,097	10	0.50
0.01	9,947,220	13	0.81

Cloud 3

N: 42,869,489 points
extents: 400 x 217 x 56 m

Voxel spacing (m)	surviving points	read time (s)	~ memory usage (GB)
0.1	2,606,613	32	0.22
0.05	7,602,382	37	0.64
0.025	19,556,071	50	1.68
0.01	37,853,465	65	3.31

Cloud 4

N: 165,155,435 points
extents: 259 x 234 x 193 m

Voxel spacing (m)	surviving points	read time (s)	~ memory usage (GB)
0.1	1,506,216	72	0.37
0.05	4,552,934	77	0.82
0.025	12,987,644	104	0.88
0.01	41,461,462	140	3.44

Conclusions

Memory usage hasn't changed
Speed improved a bit - up to ~10% in some cases, but overall it was the same

Jean-Romain commented 5 years ago

Why do you say 10% improvement. Comparing the two tables I see a > 50% improvement.

tiagodc commented 5 years ago

The time is proportional to the CPU usage... the first time I ran the tests my PC was in power management mode and only used the CPU at 10%... the second time it was 25% as it was charging, so the speed improvement was 2.5 times due to the CPU, any extra is due the usage of unordered_set instead of set...

Jean-Romain commented 5 years ago

Slightly better in computation time. No worst in memory usage. Ok I merge.

r-lidar / rlas