This is a work mainly focused on optimizing KFusion algorithm for FPGAs using OpenCL. The objective is to see its performance and check if it can be used in an heterogeneous system having also a CPU and a GPU. This research has been conducted with the following workstation:
- Intel i7-6700K
- NVIDIA Titan X - 12 GB - PCIe 3.0 x16
- Terasic DE5-Net Stratix V Altera FPGA - 4 GB - PCIe 3.0 x8
- Platform global memory: 64 GB RAM
Additionally, in order to compare the performance of different environments, some experiments have been run in a DE1SoC platform:
- ARM A9 Dual Core
- On-chip Terasic Cyclone V Altera FPGA - 64 MB
- Platform global memory: 1 GB RAM
This work is based on the SLAMBench benchmark. The framework has been enhanced so that finer-grain metrics (times, bandwidth, etc) can be obtained. Moreover, additional Python scripts have been developed to ease the use of the benchmark and provide more analytics to better understand the code performance.
The following command will run the OpenCL version 20 times with ICL-NUIM dataset 2:
python makeRun.py 20 2 opencl
The output results are the average of all runs.
Versions have been developed under specific branches:
original*
branches contains the original implementation.opt*
branches contains explored optimizations targetting the FPGA.*FPGAde1soc*
and*FPGAde5net*
refer to the target FPGA, whereas*GPU*
refers to the GPU.- kernel names in the branch name means that kernel is run in the target device whereas the other kernels are run in the host.
- Explored optimizations are also appended to the branch name.
E.g.: optFPGAde5net_reduce_fixedPoint_seqXwi_memAccessPattern
has the version to run in the DE5-Net FPGA the reduce kernel optimized with:
- Using fixed point representation and arithmetics instead of floating point.
- Variable number of work items.
- Enhanced memory access pattern.
NOTE: These branches only have kfusion/src/opencl/* files up to date. For every other single file master branch is up to date.
SLAMBench Release Candidate 1.2
Copyright (c) 2014 University of Edinburgh, Imperial College, University of Manchester.
Developed in the PAMELA project, EPSRC Programme Grant EP/K008730/1
- A SLAM performance benchmark that combines a framework for quantifying quality-of-result with instrumentation of execution time and energy consumption.
- It contains a KinectFusion (http://research.microsoft.com/pubs/155378/ismar2011.pdf) implementation in C++, OpenMP, OpenCL and CUDA (inspired by https://github.com/GerhardR).
- It offers a platform for a broad spectrum of future research in jointly exploring the design space of algorithmic and implementation-level optimisations.
- Target desktop, laptop, mobile and embedded platforms. Tested on Ubuntu, OS X and Android (on Android only the benchmark application has been ported, see later).
If you use SLAMBench in scientific publications, we would appreciate citations to the following paper (http://arxiv.org/abs/1410.2167):
L. Nardi, B. Bodin, M. Z. Zia, J. Mawer, A. Nisbet, P. H. J. Kelly, A. J. Davison, M. Luján, M. F. P. O’Boyle, G. Riley, N. Topham, and S. Furber. Introducing SLAMBench, a performance and accuracy benchmarking methodology for SLAM. In IEEE Intl. Conf. on Robotics and Automation (ICRA), May 2015. arXiv:1410.2167.
Bibtex entry:
#!latex
@inproceedings{Nardi2015,
author={Nardi, Luigi and Bodin, Bruno and Zia, M. Zeeshan and Mawer, John and Nisbet, Andy and Kelly, Paul H. J. and Davison, Andrew J. and Luj\'an, Mikel and O'Boyle, Michael F. P. and Riley, Graham and Topham, Nigel and Furber, Steve},
title = "{Introducing SLAMBench, a performance and accuracy benchmarking methodology for SLAM}",
booktitle = "{IEEE Intl. Conf. on Robotics and Automation (ICRA)}",
year = {2015},
month = {May},
NOTE = {arXiv:1410.2167}
}
If you want to set up for OS X go to the relevant section.
- TooN: maths library.
- CMake 2.8+ : building tool.
#!shell
git clone git://github.com/edrosten/TooN.git
cd TooN
./configure
sudo make install
sudo apt-get install cmake
+(with Ubuntu, you might need to install the build-essential package using sudo apt-get update && sudo apt-get install build-essential
)
-
OpenMP : for the OpenMP version
-
CUDA : for the CUDA version
-
OpenCL : for the OpenCL version (OpenCL 1.1 or greater)
-
OpenGL / GLUT : used by the graphical interface
-
OpenNI : for the live mode, and for
oni2raw
tool (which convert an OpenNI file to the SLAMBench internal format) -
Freenect Drivers : In order to use the live mode.
-
PkgConfig / Qt5 (using OpenGL) : used by the Qt graphical interface (not fully required to get a graphical interface)
-
Python (numpy) : use by benchmarking scripts (
mean
,max
,min
functions)
On ARM board, the default release of Qt5 was compile using OpenEGL, to use the Qt interface, you will have to compile Qt :
#!
cd ~
wget http://download.qt-project.org/official_releases/qt/5.2/5.2.1/single/qt-everywhere-opensource-src-5.2.1.tar.gz
tar xzf qt-everywhere-opensource-src-5.2.1.tar.gz
cd ~/qt-everywhere-opensource-src-5.2.1
./configure -prefix ~/.local/qt/ -no-compile-examples -confirm-license -release -nomake tests -nomake examples
make
make install
Then simply build doing:
#!
make
To use qt, if you compile the source as explained above, you should need to specify the Qt install dir :
#!
CMAKE_PREFIX_PATH=~/.local/qt/ make
The compilation builds 3 application modes which act like wrappers (the kernels are the same for all applications):
- benchmark: terminal user interface mode for benchmarking purposes
- main: GLUT GUI visualisation mode
- qmain: Qt GUI visualisation mode
Each application mode is also declined in 4 different builds/implementations:
- C++ (./build/kfusion/kfusion-main-cpp)
- OpenMP (./build/kfusion/kfusion-main-openmp)
- OpenCL (./build/kfusion/kfusion-main-opencl)
- CUDA (./build/kfusion/kfusion-main-cuda)
The Makefile will automatically build the executable with satisfied dependencies, e.g. the OpenCL application version will be built only if the OpenCL tool kit is available on the system and so on for CUDA, OpenMP and Qt.
All application modes and implementations share the same set of arguments:
#!plain
-c (--compute-size-ratio) : default is 1 (same size)
-d (--dump-volume) <filename> : Output volume file
-f (--fps) : default is 0
-i (--input-file) <filename> : Input camera file
-k (--camera) : default is defined by input
-l (--icp-threshold) : default is 1e-05
-o (--log-file) <filename> : default is stdout
-m (--mu) : default is 0.1
-p (--init-pose) : default is 0.5,0.5,0
-q (--no-gui) : disable any gui used by the executable
-r (--integration-rate) : default is 1
-s (--volume-size) : default is 2,2,2
-t (--tracking-rate) : default is 1
-v (--volume-resolution) : default is 256,256,256
-y (--pyramid-levels) : default is 10,5,4
-z (--rendering-rate) : default is 4
SLAMBench supports several input streams (how to use these inputs is described later):
- ICL-NUIM dataset (http://www.doc.ic.ac.uk/~ahanda/VaFRIC/iclnuim.html)
- RGB-D camera like Microsoft Kinect or other PrimeSense cameras using the OpenNI interface
- OpenNI pre-recorded file
- Raw format
Use this mode for benchmarking proposes. The output is:
- frame : ID number of the current frame
- acquisition : input data acquisition elapsed time (file reading)
- preprocessing : pre-processing elapsed time (includes kernels mm2meters bilateralFilter)
- tracking : tracking elapsed time (includes kernels halfSample, depth2vertex vertex2normal, track, reduce and solve)
- integration : integration elapsed time (includes kernel integrate)
- raycast : raycast elapsed time (include kernel raycast)
- rendering : rendering elapsed time (includes kernels renderDepth renderTrack and renderVolume)
- computation : pre-processing + tracking + integration + raycast. This is the total elapsed time for processing a frame but not including the acquiring and the visualisation kernels
- total : computation + acquisition + rendering. This is the total elapsed time for processing one frame (including the acquiring and the visualisation kernels)
- X,Y,Z : estimation of the camera position (tracking result)
- tracked : this boolean indicates if for the current frame we have not lost the tracking of the camera (1 = tracking, 0 = tracking lost)
- integrated : this boolean indicates if the integration step occurs for the current frame (depending of the tracking result and of the integration rate)
SLAMBench provides an interface to the ICL-NUIM dataset. This enables the accuracy evaluation on a SLAM implementation via the ICL-NUIM ground truth. ICL-NUIM provides 4 trajectories, we pick trajectory 2 and show how to use the dataset (for the download of each trajectory we recommend 2 GB of space available on the system):
#!plain
mkdir living_room_traj2_loop
cd living_room_traj2_loop
wget http://www.doc.ic.ac.uk/~ahanda/living_room_traj2_loop.tgz
tar xzf living_room_traj2_loop.tgz
cd ..
You can use the ICL-NUIM dataset in its native format or in a RAW format (with the latter acquiring speed increases). RAW is the format to be used for benchmarking purposes, to generate the RAW file:
#!plain
./build/kfusion/thirdparty/scene2raw living_room_traj2_loop living_room_traj2_loop.raw
Run SLAMBench:
#!plain
./build/kfusion/kfusion-benchmark-cuda -i living_room_traj2_loop.raw -s 4.8 -p 0.34,0.5,0.24 -z 4 -c 2 -r 1 -k 481.2,480,320,240 > benchmark.log
You can replace cuda by openmp, opencl or cpp.
In order to check the accuracy of your tracking compared to the ground truth trajectory, first download the ground truth trajectory file:
#!plain
wget http://www.doc.ic.ac.uk/~ahanda/VaFRIC/livingRoom2.gt.freiburg
And then use the following tool:
#!plain
./kfusion/thirdparty/checkPos.py benchmark.log livingRoom2.gt.freiburg
Get slambench data.
slambench result : 882 positions.
NUIM result : 880 positions.
Working position is : 880
Runtimes are in seconds and the absolute trajectory error (ATE) is in meters.
The ATE measure accuracy, check this number to see how precise your computation is.
Acceptable values are in the range of few centimeters.
tracking Min : 0.005833 Max : 0.046185 Mean : 0.020473 Total : 18.05721755
integration Min : 0.003629 Max : 0.041839 Mean : 0.021960 Total : 19.36882799
rendering Min : 0.018242 Max : 0.022144 Mean : 0.018486 Total : 16.30500085
preprocessing Min : 0.000633 Max : 0.002064 Mean : 0.000719 Total : 0.63432674
computation Min : 0.010156 Max : 0.075946 Mean : 0.043152 Total : 38.06037227
total Min : 0.028520 Max : 0.094302 Mean : 0.061724 Total : 54.44080794
ATE Min : 0.000000 Max : 0.044235 Mean : 0.018392 Total : 16.18503741
acquisition Min : 0.000069 Max : 0.000404 Mean : 0.000086 Total : 0.07543481
This is a GUI mode which internally uses GLUT for the visualisation step. We do not suggest to use this mode for benchmarking purposes because the visualisation step can interfere with the computation elapsed time (see http://arxiv.org/abs/1410.2167 for more information).
An example of use of the main application is:
#!plain
./build/kfusion/kfusion-main-cpp -i ~/Downloads/living_room_traj2_loop/
To run the live mode (using a Depth Sensor), you need to connect the Sensor, and to install the proper drivers. Here is the procedure with Fedora 24 :
# root user
yum install libudev-devel
Then as normal user, you will compile OpenNI2:
git clone https://github.com/occipital/OpenNI2.git OpenNI2dev
cd OpenNI2dev
sed -i.bak "s/.javaDocExe, .-d., .java../[javaDocExe, '-d', 'java', '-Xdoclint:none']/" Source/Documentation/Runme.py
cd Packaging
./ReleaseVersion.py x64
This command output finishes by Done
. Then you need to install OpenNI2 :
cd Final
mkdir -p ~/.local/OpenNI2/
tar xf OpenNI-Linux-x64-2.2.tar.bz2 -C ~/.local/OpenNI2
cd ~/.local/OpenNI2/OpenNI-Linux-x64-2.2/
export OPENNI2_INCLUDE=/home/toky/.local/OpenNI2/OpenNI-Linux-x64-2.2/Include
export OPENNI2_REDIST=/home/toky/.local/OpenNI2/OpenNI-Linux-x64-2.2/Redist
Possible error :
[toky@localhost Packaging]$ ./ReleaseVersion.py x64
Creating installer for OpenNI 2.2 x64
Traceback (most recent call last):
File "./ReleaseVersion.py", line 170, in <module>
subprocess.check_call(['make', '-C', '../', '-j' + calc_jobs_number(), 'PLATFORM=' + plat, 'release'], stdout=buildLog, stderr=buildLog)
File "/usr/lib64/python2.7/subprocess.py", line 541, in check_call
raise CalledProcessError(retcode, cmd)
subprocess.CalledProcessError: Command '['make', '-C', '../', '-j2', 'PLATFORM=x64', 'release']' returned non-zero exit status 2
Our advice to fix this error is to manually run the command, to see the real problem (in this case we have seen libudev.h was missing) :
toky@localhost Packaging]$ make -C .. PLATFORM=x64 release
make: Entering directory '/home/toky/work/pamela/slambench/OpenNI2'
make -C ThirdParty/PSCommon/XnLib/Source
make[1]: Entering directory '/home/toky/work/pamela/slambench/OpenNI2/ThirdParty/PSCommon/XnLib/Source'
g++ -MD -MP -MT "./../Bin/Intermediate/x64-Release/libXnLib.a/XnLinuxUSB.d ../Bin/Intermediate/x64-Release/libXnLib.a/XnLinuxUSB.o" -c -msse3 -Wall -O2 -DNDEBUG -I../Include -fPIC -fvisibility=hidden -Werror -o ../Bin/Intermediate/x64-Release/libXnLib.a/XnLinuxUSB.o Linux/XnLinuxUSB.cpp
Linux/XnLinuxUSB.cpp:40:21: fatal error: libudev.h: No such file or directory
#include <libudev.h>
^
compilation terminated.
../../BuildSystem/CommonCppMakefile:130: recipe for target '../Bin/Intermediate/x64-Release/libXnLib.a/XnLinuxUSB.o' failed
make[1]: *** [../Bin/Intermediate/x64-Release/libXnLib.a/XnLinuxUSB.o] Error 1
make[1]: Leaving directory '/home/toky/work/pamela/slambench/OpenNI2/ThirdParty/PSCommon/XnLib/Source'
Makefile:121: recipe for target 'ThirdParty/PSCommon/XnLib/Source' failed
make: *** [ThirdParty/PSCommon/XnLib/Source] Error 2
make: Leaving directory '/home/toky/work/pamela/slambench/OpenNI2'
To use a Kinect sensor, you will need libfreenect :
git clone https://github.com/OpenKinect/libfreenect.git
cd libfreenect
mkdir build
cd build
cmake .. -DBUILD_OPENNI2_DRIVER=ON
make
cp -L lib/OpenNI2-FreenectDriver/libFreenectDriver.so ~/.local/OpenNI2/OpenNI-Linux-x64-2.2/Redist/OpenNI2/Drivers/
This is a GUI mode which internally uses Qt for the visualisation step. We do not suggest to use this mode for benchmarking purposes because the visualisation step can interfere with the computation elapsed time (see http://arxiv.org/abs/1410.2167 for more information).
An example of use of the mainQt application is:
#!plain
./build/kfusion/kfusion-qt-cpp -i ~/Downloads/living_room_traj2_loop/
#!plain
KERNEL_TIMINGS=1 ./build/kfusion/kfusion-benchmark-cpp -s 4.8 -p 0.34,0.5,0.24 -z 4 -c 2 -r 1 -k 481.2,480,320,240 -i living_room_traj2_loop.raw -o benchmark.2.cpp.log 2> kernels.2.cpp.log
#!plain
KERNEL_TIMINGS=1 ./build/kfusion/kfusion-benchmark-openmp -s 4.8 -p 0.34,0.5,0.24 -z 4 -c 2 -r 1 -k 481.2,480,320,240 -i living_room_traj2_loop.raw -o benchmark.2.openmp.log 2> kernels.2.openmp.log
It is possible to profile OpenCL kernels using an OpenCL wrapper from the thirdparty folder :
#!plain
LD_PRELOAD=./build/kfusion/thirdparty/liboclwrapper.so ./build/kfusion/kfusion-benchmark-opencl -s 4.8 -p 0.34,0.5,0.24 -z 4 -c 2 -r 1 -k 481.2,480,320,240 -i living_room_traj2_loop.raw -o benchmark.2.opencl.log 2> kernels.2.opencl.log
The CUDA profiling takes advantage of the NVIDIA nvprof profiling tool.
#!plain
nvprof --print-gpu-trace ./build/kfusion/kfusion-benchmark-cuda -s 4.8 -p 0.34,0.5,0.24 -z 4 -c 2 -r 1 -k 481.2,480,320,240 -i living_room_traj2_loop.raw -o benchmark.2.cuda.log 2> tmpkernels.2.cuda.log || true
cat tmpkernels.2.cuda.log | kfusion/thirdparty/nvprof2log.py > kernels.2.cuda.log
When using the ICL-NUIM dataset, it is possible to generate the timings using only one command. If the living room trajectory files (raw + trajectory ground truth) are not in the directory they will be automatically downloaded.
In the following command, we use trajectory 2 and we generate the timings for the OpenCL version only:
make 2.opencl.log
In order to use all the available languages and for all the available trajectories do the following:
make 0.cpp.log 0.opencl.log 0.openmp.log 0.cuda.log
make 1.cpp.log 1.opencl.log 1.openmp.log 1.cuda.log
make 2.cpp.log 2.opencl.log 2.openmp.log 2.cuda.log
make 3.cpp.log 3.opencl.log 3.openmp.log 3.cuda.log
These parameters are also present in the Makefile and can be used with the command above, see Automatic timings section.
Trajectory 0 : offset = 0.34,0.5,0.24, voxel grid size = 5.0m (max ATE = 0.1m, mean ATE = 0.01m)
Trajectory 1 : offset = 0.485,0.5,0.55, voxel grid size = 5.0m (max ATE = 0.06m, mean ATE = 0.028m)
Trajectory 2 : offset = 0.34,0.5,0.24, voxel grid size = 4.8m (max ATE = 0.044m, mean ATE = 0.02m)
Trajectory 3 : offset = 0.2685,0.5,0.4, voxel grid size = 5.0m (max ATE = 0.292m, mean ATE = 0.117m)
#!Plain
SlamBench
├── (build) : will content the compilation and test result.
├── cmake : cmake module file
└── kfusion : kfusion test case.
├── include : common files (including the tested kernel prototypes)
├── src
│ ├── cpp : C++/OpenMP implementation
│ ├── opencl : OpenCL implementation
│ ├── cuda : CUDA implementation
└── thridparty : Includes several tools use by Makefile and 3rd party headers.
###Power measurement ### Currently power measurement is only implemented on Hardkernel boards implementing power monitoring using on board sensors, ODROID-XUE and ODROID-XU3. When executing on these platforms a file power.rpt is produced at the end of each run, this will contain a frame by frame analysis of mean power and time for each frame. The power is separated between, the Cortex-A7, Cortex-A15, GPU and DRAM.
Running kfusion/thirdparty/processPowerRpt <power.rpt> will produce total energy used by each resource during the processing of the sequence.
In the Qt interface power is monitored and graphed on a frame by frame basis, the graphing will only be visible on a suitably equipped board. The power monitor, and indeed all statistics logging, is on a per sequence basis i.e. if you open/restart a scene/file the logging will restart. The statistics for a run can be saved by selecting File->Save power from the main menu when the run has been completed, statistics are only recorded when frames are processing, pausing the sequence should not impact on the final statistics.
In the future we propose to add support for power estimation using on chip counters where available, this should be included in a future release.
Warning: SLAMBench is widely tested and fully supported on Ubuntu. The OS X version may result instable. We reckon to use the Ubuntu version.
Install Homebrew:
#!bash
ruby -e "$(curl -fsSLhttps://raw.githubusercontent.com/Homebrew/install/master/install)"
Install and change the default to gcc and g++ (only clang/LLVM available otherwise):
#!bash
brew tap homebrew/versions
brew install gcc48
sudo mv /usr/bin/gcc /usr/bin/gccORIG
sudo ln -s /usr/local/bin/gcc-4.8 /usr/bin/gcc
sudo mv /usr/bin/g++ /usr/bin/g++ORIG
sudo ln -s /usr/local/bin/g++-4.8 /usr/bin/g++
sudo mv /usr/bin/c++ /usr/bin/c++ORIG
sudo ln -s /usr/local/bin/g++-4.8 /usr/bin/c++
OpenCL is already installed out-of-the-box no need to install. Install CUDA ( guide here: http://docs.nvidia.com/cuda/cuda-getting-started-guide-for-mac-os-x/#axzz3L7QjZMEC)
Install Qt:
#!bash
brew install qt
Add to your ~/.bashrc:
#!bash
export PATH=/usr/local/opt/qt5/bin/:$PATH
Install OpenNI and dependencies (needs MacPorts already installed https://www.macports.org/install.php): Download OpenNI here and try to run the Samples/Bin/SimpleViewer program to check if the camera is working.
#!bash
sudo port install libtool
sudo port install libusb + universal
cd OpenNI
./install.sh
In SLAMBench, modify cmake/FindOpenNI.cmake adding to the lib path:
/Users/lnardi/sw/OpenNI-MacOSX-x64-2.2/Samples/Bin
And to the include path:
/Users/lnardi/sw/OpenNI-MacOSX-x64-2.2/Include
Add to your ~/.bashrc:
#!bash
export DYLD_LIBRARY_PATH=/Users/lnardi/sw/OpenNI-MacOSX-x64-2.2/Samples/Bin:$DYLD_LIBRARY_PATH
##Known Issues ##
- ** Failure to track using OpenCL on AMD ** -Some issues have been reported on AMD platforms, we will look into this
- Visualisation using QT - may be offset on some platforms - notably ARM
- **Build issues using QT on ARM ** - Visualisation requires opengl but Qt on ARM is often built using GLES, including packages obtained from distribution repo. Building from source with opengl set to desktop resolves this.
- ** Frame rates on QT GUI appear optimistic** - The rate shown in the status bar is by default the computation time to process the frame and render any output, it excludes the time take by the QT interface to display the rendered images and acquire frame
- ** performance difference between CUDA/OpenCL** - This is a known issue that we are investigating. It's mainly cause by a difference of global work-group size between the both version and a major slowdown of CUDA in the rendering kernels is cause by the use of float3 instead of float4 which result by an alignment issue. this alignment issue doesn't appear in OpenCL as cl_float3 are the same as cl_float4.
- ** CUDA nvprof slows down the performance on some platforms** - the nvprof instrumentation has a 2x slowdown on MAC OS for the high-level KFusion building blocks. So if we run using make 2.cuda.log we will not measure the maximum speed of the machine for the high-level building blocks. It is questionable then if we should keep measuring the CUDA high-level and low-level performance at the same time or in order to be more accurate it is better to run the two measurements in two separate runs.
- ** OS X version has not been widely tested **
Release candidate 1.2 (To be define)
- Bugfix : remove extra line of log.
- ...
Release candidate 1.1 (17 Mar 2015)
- Bugfix : Move bilateralFilterKernel from preprocessing to tracking
- Bugfix : Wrong interpretation of ICP Threshold parameter.
- Esthetic : Uniformisation of HalfSampleRobustImage kernel
- Performance : Change float3 to float4 for the rendering kernels (No effect on OpenCL, but high performance improvement with CUDA)
- Performance : Add a dedicated buffer for the OpenCL rendering
- Feature : Add OSX support
Release candidate 1.0 (12 Nov 2014)
- First public release