OpenCV 3.

1 cross compilation on Linux

A15
Instructions:
1. Required host machine: Ubuntu 10.04 or 12.04 or 14.04

Note: to copy from terminal – shift + ctrl + c and to paste into terminal – shift + ctrl + v
2. Let’s make a folder called ‘ti’ in the home folder and keep the OpenCV package and Vision SDK
package there
a. cd ~
b. mkdir ti
3. Download and install Vision SDK 02.11
4. Please follow the steps mentioned in VisionSDK_LinuxUserGuide.pdf in
<vision_sdk_install_path>/VISION_SDK_02_08_00_00/vision_sdk/linux/docs folder
5. It is expected that the user has installed the A15 compiler as mentioned in section 2.2.1 in the
user guide
6. So, the linaro toolchain should be found inside
~/ti/VISION_SDK_02_08_00_00/ti_components/os_tools/linux/linaro
7. OpenCV setup + applying patch
a. cd ~/ti
b. mkdir opencv
c. cd opencv
d. clone tiopencv repository using
i. git clone https://github.com/opencv/opencv.git
e. Go to the tag ticv3.1_00.04.02.00
f. Create a build directory called ‘build_arm’
8. Arm-toolchain file
a. The Linux ARM cmake toolchain is part of the patch.
b. It should be found inside the <opencv_path>/platforms/generic/ named ‘arm-
gnueabi.toolchain_vsdk.cmake’
c. The build depends on the following components:
i. GCC
ii. OpenCL (part of VSDK Linux targetfs)
iii. VXLIB
iv. DSP CGT 8.1.0
d. Go through the toolchain and edit the appropriate paths to point to the appropriate
modules path mentioned earlier

Note: Let’s keep the FLOAT_ABI_SUFFIX as ‘hf’ only and not disturb it unless required

Texas Instruments NDA restrictions Page 1

 9. Install 32-bit libraries if the host machine is 64-bit and the downloaded toolchain is 32-bit
a. Cmd: sudo apt-get install ia32-libs
10. Download cmake
a. Open Terminal – ctrl + alt + t
b. Cmd: sudo apt-get install cmake
c. Cmd: sudo apt-get install cmake-qt-gui
11. Configuring OpenCV project
a. Open cmake gui
i. Cmd: cmake-qt-gui
b. Choose source path: /home/<username>/ti/opencv/tiopencv
c. Choose build path: /home/<username>/ti/opencv/build_arm
d. Press configure -> select toolchain file -> location:
~/ti/opencv/tiopencv/platforms/linux/arm-gnueabi.toolchain_vsdk.cmake (~ stands for
/home/<username>)
e. Select reqd. modules as follows:

Texas Instruments NDA restrictions Page 2

Texas Instruments NDA restrictions Page 3
Texas Instruments NDA restrictions Page 4
Texas Instruments NDA restrictions Page 5
Texas Instruments NDA restrictions Page 6
Texas Instruments NDA restrictions Page 7
 f. OpenMP and pthread can be enabled if multi-core processing is desired

Note: all the opencv kernels may not be parallelized

g. Click configure
h. Click generate
i. Close cmake
j. Cmd: make –j4
k. Cmd: make install
12. If the above build fails, with an assembler statement saying “offset out of range”, we need the
64-bit version of the same toolchain
a. Get the toolchain package
i. Cmd: sudo gedit /etc/apt/sources.list
ii. Go to eof
iii. Add the line – deb http://cz.archive.ubuntu.com/ubuntu trusty main universe
iv. Save and close the file
v. Cmd: sudo apt-get update
b. Cmd: sudo apt-get install gcc-4.7-arm-linux-gnueabihf
c. Cmd: sudo apt-get install g++-4.7-arm-linux-gnueabihf
13. Edit the arm-toolchain file if step 12 was followed
a. Open terminal – ctrl + alt + t
b. gedit ~/ti/opencv/opencv-3.1.0/platforms/linux/arm-gnueabi.toolchain.cmake
c. Edit the following lines
i. find_program(CMAKE_C_COMPILER NAMES arm-linux-
gnueabi${FLOAT_ABI_SUFFIX}-gcc-4.7)
ii. find_program(CMAKE_CXX_COMPILER NAMES arm-linux-
gnueabi${FLOAT_ABI_SUFFIX}-g++-4.7)
iii. set(ARM_LINUX_SYSROOT /usr/arm-linux-gnueabihf PATH "ARM cross
compilation system root")
14. Go to project configuration
a. Clear cache
b. Repeat step 11(Configuring OpenCV project..)
c. Now the build should be successful
15. Installing the libraries
a. Run this command
i. Cmd: make install
b. This will install the headers and libraries in install folder inside the build_arm folder
c. The demo or test executables can be found inside the bin folder of build_arm folder

Texas Instruments NDA restrictions Page 8

List of supported modules
Below list of modules are supported for BIOS build

S. No Modules
1 calib3d
2 core
3 features2d
4 flann
5 imgcodecs
6 imgproc
7 ml
8 objdetect
9 photo
10 shape
11 stitching
12 superres
13 video
14 videostab

List of OpenCV functions accelerated using OpenCL on DSP

For performance data refer to the performance excel sheets:

1. vayu_arm_linux_opencv_test_report.xls – OpenCV tests for arm linux

2. vayu_arm_bios_opencv_test_report.xls – OpenCV tests for arm linux
3. OpenCV_offload_DSP_profiling.xlsx – OpenCV tests for DSP accelerated functions

Function Constraints Introduced in

VSDK version
cv::erode 8-bit single channel input; 8-bit single channel output; only 3x3 2.11
structuring is supported
cv::dilate 8-bit single channel input; 8-bit single channel output; only 3x3 2.11
structuring is supported
cv::GaussianBlur 8-bit single channel input; 8-bit single channel output; only 3x3 2.11

Texas Instruments NDA restrictions Page 9

 structuring is supported
cv::medianBlur 8-bit single channel input; 8-bit single channel output; only 3x3 2.11
structuring is supported
cv::LUT 8-bit single channel input 2.12
cv::MorphologyEx 8-bit single channel input; single iteration only; default anchor 2.12
only supported
cv::PyrDown 8-bit single channel input; difference in implementation 2.12
compared to opencv – the difference is in Gaussian kernel
which is

1/256 of | 1 4 6 4 1|
| 4 16 24 16 4|
| 6 24 36 24 6|
| 4 16 24 16 4|
| 1 4 6 4 1|

cv::Sobel 8-bit single channel input; only 3x3 structuring is supported; 2.12
sobel xy calculates magnitude also and output is 16-bit single
channel
cv::Resize 8-bit single channel input; only bilinear and downscaling to half 2.12
size is supported
cv::Integral 8-bit single channel input; 2.12
cv::CalcHist 8-bit single channel input; only 256 bins and range of 256 is 2.12
supported
cv::EqualizeHist 8-bit single channel input 2.12

Accelerating more VLIB/VXLIB kernels

1. Before Building new OpenCL wrappers for optimized DSP kernels
 In order to build OpenCL
o The OpenCL cross compiler, ‘clocl’ is required (this can be found inside the
ti_components directory of Vision SDK) and
o The appropriate DSP symbols
 The DSP symbols directory needs to be exported using an environment variable,
‘TARGET_ROOTDIR’
 Please follow the FeatureSpecificguidelines on OpenCX for setting up the
‘TARGET_ROOTDIR’
 Also, point to the CLOCL path in the OpenCV Linux cmake toolchain

2. Building new OpenCL wrappers for optimized DSP kernels

 The above list of OpenCV functions was accelerated using TI VXLIB C66X kernels.
 The OpenCL wrapper around these optimized DSP kernels could be found inside each
opencv module inside ‘src/ti_opencl’ folder

Texas Instruments NDA restrictions Page 10

 o i.e
 <path to tiopencv>/modules/imgproc/src/ti_opencl
 OpenCL 1.1 allows calling any C function from a target library
o In order to do this, an interface to the function is needed and
o The target library to link with
 If new OpenCL wrappers for other VXLIB kernels are going to be added to any OpenCV
module
o Create appropriate .cl file
o Edit the interface file to add prototype of the function to be called
o Then do ‘cmake’ and make
o The OpenCV build system is improved to automatically build all opencl files inside
the ‘ti_opencl’ directory
 In order to link with new DSP libraries such as VLIB
o Add appropriate interface files in ti_opencl directory
o Edit ‘OpenCVModule.cmake’ inside <path to tiopencv>/cmake
 Search for the string ‘CLOCL_CMD
 It could be seen that a command is formed to build the OpenCL file
 Add the new library path in the command to link with it

*Note: In order to build the TI OpenCL files, pleae run ‘cmake’ which builds it.

3. Utilizing the OpenCL wrappers

 After the TI OpenCL files are build the binary is stored in a character array in a correspding
header inside the <opencv_build_dir_path>/modules/<module_name>/ti_opencl/<header>
 Include the header in the appropriate source file.
o E.g. See morph.cpp
o Add an equivalent function to call TI OpenCL file e.g. ‘ticl_morph’
o Acquire the OpenCL device using ‘ocl::Device::getDefault’
o Create a ‘ocl::ProgramBinary’ object from the binary string
o Set appropriate global and local sizes
o Create a ‘ocl::Kernel’ object from ‘ocl::ProgramBinary ‘ object
o Set the kernel arguments
o Finally use the method ‘run’ of ‘ocl::Kernel’ object to offload and run the kernel
 Make sure that the ‘ticl_<kernel_name>’ function is called inside the appropriate OpenCV
C++ function.

Texas Instruments NDA restrictions Page 11