Improving performance for ARM7-based platforms #18
Conversation
dkargin
commented
Dec 20, 2016
dkargin
commented
- most calculations use float instead of double (for NEON instruction set)
- column-order calculation, less 'atan2' calls
- added tests for performance comparison. 20x faster so far (tests were not so reliable)
- added test node to compare both scans from old and new processor
- most calculations use float instead of double - column-order calculation, less 'atan2' calls - added tests for perfomance comparison. 20x faster so far (tests were not so reialble) - added test node to compare both scans from old and new processor
| @@ -64,8 +64,45 @@ namespace depthimage_to_laserscan | |||
| * | |||
| */ | |||
| sensor_msgs::LaserScanPtr convert_msg(const sensor_msgs::ImageConstPtr& depth_msg, | |||
There was a problem hiding this comment.
inline needed due to implementation in the header?
There was a problem hiding this comment.
Moved implementation to cpp. Only template part is left in header
| * @return sensor_msgs::LaserScanPtr for the center row(s) of the depth image. | ||
| * | ||
| */ | ||
| sensor_msgs::LaserScanPtr convert_msg_f(const sensor_msgs::ImageConstPtr& depth_msg, |
There was a problem hiding this comment.
Inline here too. But I'd suggest that we don't need to add the wrapper on the allocation free conversion. Passing in the message is a cleaner API. And for the double one, it's important to keep it available for backwards compatibility, but since this is a new API we don't need to add the parallel interface.
| @@ -102,7 +118,7 @@ sensor_msgs::LaserScanPtr DepthImageToLaserScan::convert_msg(const sensor_msgs:: | |||
| double angle_min = -angle_between_rays(center_ray, right_ray); // Negative because the laserscan message expects an opposite rotation of that from the depth image | |||
|
|
|||
| // Fill in laserscan message | |||
| sensor_msgs::LaserScanPtr scan_msg(new sensor_msgs::LaserScan()); | |||
| //sensor_msgs::LaserScanPtr scan_msg(new sensor_msgs::LaserScan()); | |||
There was a problem hiding this comment.
Please delete the old code instead of just commenting it.
| @@ -184,7 +245,8 @@ namespace depthimage_to_laserscan | |||
| int offset = (int)(cam_model.cy()-scan_height/2); | |||
| depth_row += offset*row_step; // Offset to center of image | |||
|
|
|||
| for(int v = offset; v < offset+scan_height_; v++, depth_row += row_step){ | |||
| for(int v = offset; v < offset+scan_height_; v++, depth_row += row_step) | |||
| { | |||
There was a problem hiding this comment.
Please avoid whitespace changes only to keep merges as easy as possible. I won't object to the newlines being inserted, but there are a few other lines with this rewrapping of unchanged lines which makes the review harder.
