CUDA performance

[geomat0101]

FYI, here's the examples rendered with a CUDA-enabled dlib. Most of these are going from a couple seconds down to < 0.02 seconds. The one with all the bald dudes (25 faces) went from about 20 seconds to < 0.6!

Note that there is noticeable startup cost factored into the first one rendered. We need to do our best to ensure that our face-analyzer stays active and primed so we don't eat that startup cost for every image (or batch of) that we want to process.

(mediamgr) mdg@ftl:~/src/mediamgr/python/examples$ ./cnn_face_detector.py mmod_human_face_detector.dat faces/*jpg | grep "Num\|render"
Number of faces detected: 7
rendering time: datetime.timedelta(seconds=1, microseconds=531932)
Number of faces detected: 2
rendering time: datetime.timedelta(microseconds=15500)
Number of faces detected: 3
rendering time: datetime.timedelta(microseconds=13331)
Number of faces detected: 4
rendering time: datetime.timedelta(microseconds=13535)
Number of faces detected: 6
rendering time: datetime.timedelta(microseconds=12694)
Number of faces detected: 3
rendering time: datetime.timedelta(microseconds=15377)
Number of faces detected: 7
rendering time: datetime.timedelta(microseconds=52650)
Number of faces detected: 7
rendering time: datetime.timedelta(microseconds=13121)
Number of faces detected: 2
rendering time: datetime.timedelta(microseconds=16849)
Number of faces detected: 25
rendering time: datetime.timedelta(microseconds=597486)
Number of faces detected: 0
rendering time: datetime.timedelta(microseconds=76118)
Number of faces detected: 1
rendering time: datetime.timedelta(microseconds=49817)