ga_points_downsampler

ros package points downsampler with cupoch.

optimized with jetson.

performance in jetson

110k points

tested in jetson xavier, jetpack 4.5.1

downsample a pointcloud of 119978 points with res=1.0 to 3440 points.

the bandwidth is 14.4MB/s with 10*1.44MB

• cuda-pcl

7ms

• pcl

6ms

• cupoch

4ms

build

performance tunning with jetson

NOTE: You should set power management policy for cpu and gpu to get the best performance of cudaMemCpy.

cpu frequency scaling

refer to this guide from stackoverflow to disable cpu scaling.

sudo apt-get install cpufrequtils
# to get the highest freq
sudo cpufreq-set -g performance

# to get the default mode powersave
sudo cpufreq-set -g powersave

gpu frequency scaling

refer to this guide for more details.

# If not, please set CONFIG_DEVFREQ_GOV_PERFORMANCE=y and update the kernel
zcat /proc/config.gz | grep CONFIG_DEVFREQ_GOV_PERFORMANCE
# Run below command on board as root to keep GPU alive (not suspend)
echo 0 > /sys/devices/17000000.gv11b/railgate_enable
# Enable performance governor for GPU
cat /sys/class/devfreq/17000000.gv11b/available_governors
echo performance > /sys/class/devfreq/17000000.gv11b/governor
cat /sys/class/devfreq/17000000.gv11b/governor