Closed
Description
Hi there,
I did a benchmark for cv::cvtColor (in essense it loops through a set of images and does conversion BayerRG => BGR).
Run on (24 X 2112 MHz CPU s)
CPU Caches:
L1 Data 48 KiB (x12)
L1 Instruction 32 KiB (x12)
L2 Unified 2048 KiB (x12)
L3 Unified 30720 KiB (x1)
--------------------------------------------------------------------------
Benchmark Time CPU Iterations
--------------------------------------------------------------------------
BM_ConvertBayerRG8/threads:1 141345 ns 142299 ns 5600
BM_ConvertBayerRG8/threads:2 102112 ns 207203 ns 3318
BM_ConvertBayerRG8/threads:4 143496 ns 496779 ns 2988
BM_ConvertBayerRG8/threads:8 258000 ns 898438 ns 800
BM_ConvertBayerRG8EA/threads:1 169186 ns 174386 ns 4480
BM_ConvertBayerRG8EA/threads:2 129333 ns 204041 ns 3446
BM_ConvertBayerRG8EA/threads:4 175315 ns 592913 ns 2240
BM_ConvertBayerRG8EA/threads:8 287515 ns 1063756 ns 896
BM_ConvertBayerRG8VNG/threads:1 14805108 ns 14409722 ns 90
BM_ConvertBayerRG8VNG/threads:2 13547418 ns 26562500 ns 20
BM_ConvertBayerRG8VNG/threads:4 7017534 ns 25000000 ns 40
BM_ConvertBayerRG8VNG/threads:8 3459721 ns 20507813 ns 80
And BM_ConvertBayerRG8VNG results look weird to me. How it comes that running benchmark on 4, 8 threads gives lesser times than running on a single thread?
Images (2046*2046) are of 4mb in size (the result image is 12mb). The set in total is of 33 files.
Metadata
Metadata
Assignees
Labels
No labels