Skip to content

enable concat layer fuse for OCL target#11959

Merged
opencv-pushbot merged 1 commit intoopencv:3.4from
pengli:3.4
Jul 17, 2018
Merged

enable concat layer fuse for OCL target#11959
opencv-pushbot merged 1 commit intoopencv:3.4from
pengli:3.4

Conversation

@pengli
Copy link
Copy Markdown

@pengli pengli commented Jul 13, 2018 •
edited by alalek
Loading 

force_builders=Docs,linux,ocllinux,windows,ocl,macosx,oclmacosx,linuxNoOpt

@pengli pengli force-pushed the 3.4 branch 6 times, most recently from 6ee14ed to 44f895b Compare July 16, 2018 04:15
@pengli
Copy link
Copy Markdown
Author

pengli commented Jul 16, 2018

I run GoogleNet fp16 test on SKL GT2, perf time improved from 10.2ms to 9.7ms.

@alalek
Copy link
Copy Markdown
Member

alalek commented Jul 16, 2018

i7-6700 (Fedora 28 + Intel OpenCL NEO 18.26.10987) results:

Name of Test origin patch x
AlexNet::DNNTestNetwork::(DNN_BACKEND_OPENCV, DNN_TARGET_OPENCL) 14.354 14.363 1.00
AlexNet::DNNTestNetwork::(DNN_BACKEND_OPENCV, DNN_TARGET_OPENCL_FP16) 8.592 8.642 0.99
DenseNet_121::DNNTestNetwork::(DNN_BACKEND_OPENCV, DNN_TARGET_OPENCL) 55.803 62.607 0.89
DenseNet_121::DNNTestNetwork::(DNN_BACKEND_OPENCV, DNN_TARGET_OPENCL_FP16) 69.040 79.798 0.87
EAST_text_detection::DNNTestNetwork::(DNN_BACKEND_OPENCV, DNN_TARGET_OPENCL) 96.957 97.601 0.99
EAST_text_detection::DNNTestNetwork::(DNN_BACKEND_OPENCV, DNN_TARGET_OPENCL_FP16) 76.413 77.883 0.98
ENet::DNNTestNetwork::(DNN_BACKEND_OPENCV, DNN_TARGET_OPENCL) 22.353 37.204 0.60
FastNeuralStyle_eccv16::DNNTestNetwork::(DNN_BACKEND_OPENCV, DNN_TARGET_OPENCL) 99.908 100.513 0.99
GoogLeNet::DNNTestNetwork::(DNN_BACKEND_OPENCV, DNN_TARGET_OPENCL) 14.292 13.701 1.04
GoogLeNet::DNNTestNetwork::(DNN_BACKEND_OPENCV, DNN_TARGET_OPENCL_FP16) 11.287 10.734 1.05
Inception_5h::DNNTestNetwork::(DNN_BACKEND_OPENCV, DNN_TARGET_OPENCL) 16.351 15.729 1.04
Inception_5h::DNNTestNetwork::(DNN_BACKEND_OPENCV, DNN_TARGET_OPENCL_FP16) 13.506 12.917 1.05
Inception_v2_SSD_TensorFlow::DNNTestNetwork::(DNN_BACKEND_OPENCV, DNN_TARGET_OPENCL) 49.297 49.997 0.99
Inception_v2_SSD_TensorFlow::DNNTestNetwork::(DNN_BACKEND_OPENCV, DNN_TARGET_OPENCL_FP16) 53.582 54.104 0.99
MobileNet_SSD_Caffe::DNNTestNetwork::(DNN_BACKEND_OPENCV, DNN_TARGET_OPENCL) 20.860 20.909 1.00
MobileNet_SSD_Caffe::DNNTestNetwork::(DNN_BACKEND_OPENCV, DNN_TARGET_OPENCL_FP16) 17.746 17.748 1.00
MobileNet_SSD_v1_TensorFlow::DNNTestNetwork::(DNN_BACKEND_OPENCV, DNN_TARGET_OPENCL) 22.165 22.242 1.00
MobileNet_SSD_v1_TensorFlow::DNNTestNetwork::(DNN_BACKEND_OPENCV, DNN_TARGET_OPENCL_FP16) 18.711 18.483 1.01
MobileNet_SSD_v2_TensorFlow::DNNTestNetwork::(DNN_BACKEND_OPENCV, DNN_TARGET_OPENCL) 27.008 27.103 1.00
MobileNet_SSD_v2_TensorFlow::DNNTestNetwork::(DNN_BACKEND_OPENCV, DNN_TARGET_OPENCL_FP16) 26.102 26.008 1.00
OpenFace::DNNTestNetwork::(DNN_BACKEND_OPENCV, DNN_TARGET_OPENCL) 6.851 6.887 0.99
OpenFace::DNNTestNetwork::(DNN_BACKEND_OPENCV, DNN_TARGET_OPENCL_FP16) 8.321 8.358 1.00
ResNet_50::DNNTestNetwork::(DNN_BACKEND_OPENCV, DNN_TARGET_OPENCL) 30.511 30.581 1.00
ResNet_50::DNNTestNetwork::(DNN_BACKEND_OPENCV, DNN_TARGET_OPENCL_FP16) 22.046 22.229 0.99
SSD::DNNTestNetwork::(DNN_BACKEND_OPENCV, DNN_TARGET_OPENCL) 517.671 520.363 0.99
SSD::DNNTestNetwork::(DNN_BACKEND_OPENCV, DNN_TARGET_OPENCL_FP16) 344.160 343.189 1.00
SqueezeNet_v1_1::DNNTestNetwork::(DNN_BACKEND_OPENCV, DNN_TARGET_OPENCL) 7.318 6.683 1.09
SqueezeNet_v1_1::DNNTestNetwork::(DNN_BACKEND_OPENCV, DNN_TARGET_OPENCL_FP16) 5.657 5.209 1.09
YOLOv3::DNNTestNetwork::(DNN_BACKEND_OPENCV, DNN_TARGET_OPENCL) 376.247 378.513 0.99
YOLOv3::DNNTestNetwork::(DNN_BACKEND_OPENCV, DNN_TARGET_OPENCL_FP16) 323.300 326.254 0.99
opencv_face_detector::DNNTestNetwork::(DNN_BACKEND_OPENCV, DNN_TARGET_OPENCL) 20.171 20.165 1.00
opencv_face_detector::DNNTestNetwork::(DNN_BACKEND_OPENCV, DNN_TARGET_OPENCL_FP16) 23.896 24.039 0.99

Could you take a look on performance regressions for DenseNet (11%), ENet (40%) ?

@pengli
Copy link
Copy Markdown
Author

pengli commented Jul 17, 2018

@alalek , ENet and DenseNet perf regression is fixed in the update. Pls give it a try. Thanks.

@pli2-intel
enable concat layer fuse for OCL target
f0cadaa 
Signed-off-by: Li Peng <peng.li@intel.com>
alalek
alalek approved these changes Jul 17, 2018
View reviewed changes
Copy link
Copy Markdown
Member

@alalek alalek left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Well done! Thank you 👍

@opencv-pushbot opencv-pushbot merged commit f0cadaa into opencv:3.4 Jul 17, 2018
opencv-pushbot pushed a commit that referenced this pull request Jul 17, 2018
@alalek
Merge pull request #11959 from pengli:3.4
78d07e8
@alalek alalek mentioned this pull request Jul 17, 2018
Merged
@alalek alalek mentioned this pull request Sep 4, 2021
Merged
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@alalek alalek alalek approved these changes

Assignees

@alalek alalek

Labels

category: dnn category: ocl optimization

Projects

None yet

Milestone

3.4.3

Development

Successfully merging this pull request may close these issues.

4 participants

@pengli @alalek @opencv-pushbot @pli2-intel