NEON optimisation of cv::threshold() for iOS (Feature #1455)

Added by Yasuhiro Yoshimura over 13 years ago. Updated over 13 years ago.

Status:Done Start date:
Priority:Normal Due date:
Assignee:Vadim Pisarevsky % Done:

0%
Category:imgproc, video
Target version:-
Difficulty: Pull request:

Description

I implemented NEON optimisation of cv::threshold() for iOS.
I checked that patched cv::threshold() become fast
about 10x speed up on the device(iPod touch 4th).

Maybe, I think that this patch is also effective on Android.

opencv231_neon_thresh_8u.patch - NEON optimisation of cv::threshold() for iOS (6.3 kB) Yasuhiro Yoshimura, 2011-10-31 02:03 pm

Associated revisions

Revision 2dceb68a
Added by Alexander Smorkalov over 11 years ago

Merge pull request #1455 from ilya-lavrenov:ocl_test_output

History

#1
Updated by Andrey Kamaev over 13 years ago

This code can not be included into the OpenCV, because it can fail with SIGSEG after attempt to write unallocated memory.

  • Status changed from Open to Done
  • (deleted custom field) set to invalid

#2
Updated by Yasuhiro Yoshimura over 13 years ago

Replying to [comment:1 andrey.kamaev]:

This code can not be included into the OpenCV, because it can fail with SIGSEG after attempt to write unallocated memory.

Thank you for your comment.
I understand. I should add the following processing at
the beginning of this function.

if( _src.empty() || _dst.empty() ) {
return;
}

But, if "src" and "dst" Mat are NULL, roi.width and roi.width are initialized to 0.
So, unallocated memory in not accessed.

#3
Updated by Andrey Kamaev over 13 years ago

Replying to [comment:2 dandelion]:

Replying to [comment:1 andrey.kamaev]:

This code can not be included into the OpenCV, because it can fail with SIGSEG after attempt to write unallocated memory.

Thank you for your comment.
I understand. I should add the following processing at
the beginning of this function.

if( _src.empty() |@@| _dst.empty() ) {
return;
}

But, if "src" and "dst" Mat are NULL, roi.width and roi.width are initialized to 0.
So, unallocated memory in not accessed.

Empty Mats is not a real problem of your code. You are wrong in the leftovers processing (all the cycles making j+=8). Also I should note that copying SSE optimization in NEON intrinsics rarely result in good code and I think that you can make a noticeably faster version using more suitable instructions.

Also available in: Atom PDF