Fu-. 0 Posted August 16, 2023 Hi, I am challenging image judgment using OpenCV and YOLO. When I ran libmali using OpenCL, multiple threads took the form of one mutex, and the operation speed was abnormally slow. The expected speed is about 100ms per judgment, but it takes more than 10 seconds when it actually works. Is there any way to solve this? I'm using the latest Tinker2S debian image and OpenCV 4.8.0. Regards, Share this post Link to post Share on other sites
tooz 53 Posted August 16, 2023 hello @Fu-., and are you using yolov5? 1 hour ago, Fu-. said: I'm using the latest Tinker2S debian image the debian 11 os that just released? Share this post Link to post Share on other sites
Fu-. 0 Posted August 17, 2023 (edited) Hi, I'm using yolo v7. I've just noticed that the debian 11 was released. I'll try the new image and test. Regards, Edited August 17, 2023 by Fu-. Share this post Link to post Share on other sites
Fu-. 0 Posted August 17, 2023 (edited) I've tested on the latest debian 11, but it seems to look like nothing to do with debian version. The inference is too slow. Edited August 17, 2023 by Fu-. Share this post Link to post Share on other sites
tooz 53 Posted August 17, 2023 hello @Fu-. may i knnow how your environment and how you run the inference, i'll see if i can reproduce the performance issue Share this post Link to post Share on other sites
Fu-. 0 Posted August 17, 2023 Hi, tooz. I'll provide a test environment where the problem occurs. Please download from below. https://drive.google.com/file/d/1222dOTwmF18_-sgBqk_YtPRJfL-DzPN4/view?usp=drive_link Regards, Share this post Link to post Share on other sites
tooz 53 Posted August 18, 2023 hello @Fu-., i've downloaded your file and installed opencv/ libjpeg-turbo with your cmake list, however i cannot find the sample.zip. is this file missing or is it in the other zip files? Build and execute samples $ unzip sample.zip $ cd ~/sample $ make $ export OPENCV_OCL4DNN_CONFIG_PATH=~/.cache/opencv/4.x/opencl_cache/ARM--Mali-T860--1_2 $ ./sample bus.jpg Share this post Link to post Share on other sites
Fu-. 0 Posted August 18, 2023 (edited) Hello @tooz. The zip archive contains to build the sample sources and Makefile. So, extract zip to /home/linaro/, Please make here. Regards, Edited August 18, 2023 by Fu-. Share this post Link to post Share on other sites
tooz 53 Posted August 18, 2023 hello @Fu-., i was too stupid to realize that you meant the 'reproduce.zip'. i just set up the enviorment and can confirm it takes about 1 - 2 minutes to get the result.jpg. i'll let you know as soon as possible if there's any new findings. Share this post Link to post Share on other sites
Fu-. 0 Posted August 18, 2023 Hello @tooz, Thank you. I hope you can find a good solution. Regards, Share this post Link to post Share on other sites
tooz 53 Posted August 22, 2023 hello @Fu-. On 8/16/2023 at 4:12 PM, Fu-. said: The expected speed is about 100ms per judgment, but it takes more than 10 seconds when it actually works. we are checking with the soc vendor, at the same time, may i know where this ↑ reference came from? thank you Share this post Link to post Share on other sites
Fu-. 0 Posted August 24, 2023 Hello @tooz, On 8/22/2023 at 4:20 PM, tooz said: hello @Fu-. > The expected speed is about 100ms per judgment, but it takes more than 10 seconds when it actually works. we are checking with the soc vendor, at the same time, may i know where this ↑ reference came from? thank you I'm using a different YOLO onnx. This input size is 640x640. It's not sample onnx(convertimg from yolov7l), this takes about 10sec per inference. Regards, Share this post Link to post Share on other sites