Linux/AM5728: OpenCV optimization

user5339124

Part Number: AM5728

Tool/software: Linux

Hi,

we are going to using opencl to accelerate opencv based on am5728 platform, so i have designed a simple test app, the procedure as following:

1) read left and right two images from input files

2) call and execute StereoBM algorithm and get the disparity image

3) call minMaxLoc function in order to convert disparity image form CV_16S type to CV_8u type or call getMat() function to get mat format data structure for next process.

I find that if i enable opencl, and stereoBM algorithm running faster than disable opencl, but minMaxLoc() or getMat() will take nearly 2s, it's very strange for this, because if i don't enable opencl which only take 0.1ms, i think since umat is a TAPI, so host and device all have access rights, should no data movement or transfer, but the result show me my thinking is wrong.

The test app and log file please see the attachments!

Is there any OpenCV/OpenCL expert give me a detail explanation, I am stacking at this problem for a long time.

over 5 years ago

0 Margarita Gashova over 5 years ago

TI__Guru* 80190 points

Hello,

I do not see any attachments.

BR
Margarita

0 user5339124 over 5 years ago in reply to Margarita Gashova

Prodigy 40 points

Sorry Margarita,

Thanks for your quick response, Maybe it's browser problem, i upload my test app again, Please help me out./cfs-file/__key/communityserver-discussions-components-files/791/StereoBM_5F00_UMat2Mat_5F00_2018_2D00_05_2D00_22_5F00_ForTI.rar

Howard

0 user5339124 over 5 years ago

Prodigy 40 points

Part Number: AM5728

Tool/software: Linux

Hi,

Recently i encounter some problems using OpenCV dispatch OpenCL, So i based on TI Wiki link http://processors.wiki.ti.com/index.php/OpenCV designed two simple test application, which based on stereoBM algorithm, find there is a obviously performance difference with OpenCV app dispatch OpenCL using OpenCV-OpenCl classes (Option A for simple) and using standard OpenCL dispatch (Option B for simple).

1. Using Option A, stereoBM algorithm execute only take 30ms, but the minMaxLoc function nearly take 2s.

StereoBM_OptionA tdiff=30.513989 ms

minMaxLoc tdiff=1968.670074 ms; Min disp: -16.000000 Max value: 830.000000

2. Using Option B, stereoBM algorithm execute will take nearly 2s, but the minMaxLoc function nearly 1ms.

StereoBM_OptionB tdiff=1965.977616 ms

minMaxLoc tdiff=0.984459 ms;

Whether option A or option B, the total time is almost the same( about 2s ), details please see the attachments(Test app and log file/cfs-file/__key/communityserver-discussions-components-files/791/OpenCVTestBasedOnStereoBM_5F00_2018_2D00_05_2D00_23.rar). I can't understand why, so i post the question and hope experts to answer it for me, Thanks in advance.

Howard

0 user5339124 over 5 years ago in reply to user5339124

Prodigy 40 points

Hi,

This series of problem have sucessfully resolved under the TI local FAE Thomas Yang's supports, He give me the proper answer as following:

OpenCV/OpenCL is implemented using async (“lazy”) dispatch to DSP.
That means when kernel on DSP is invoked, control typically immediately comes back to ARM code. This allows parallel execution between DSP and ARM.
But, when data dependency criteria is detected (i.e. argument of next operation is output of previously invoked DSP kernel), there is waiting for end of DSP operation (creating impression of slow variable access).

According to above tips, i change the opencl dispatch from async mode into sync mode, this is to say, After change "k.run(2, globalThreads, NULL, false);" into "k.run(2, globalThreads, NULL, true);" , All test app's output results are reasonable, in fact, the stereoBM alogrithm exhausted about 2s at current, Next i need to optimize the stereoBM algorithm.

0 Margarita Gashova over 5 years ago in reply to user5339124

TI__Guru* 80190 points

Hello,

Thank you for sharing the solution.

BR
Margarita

Processors

Processors forum

Linux/AM5728: OpenCV optimization