Rockchip RK3588 – OpenCL环境搭建

在上一节《Rockchip RK3588 - 基于Qt的视频监控和控制系统 》,我们介绍了实时监控的实现,在实时监控中我们需要将分辨率为1920x1080的图像缩放为指定窗口大小的图像,当采样帧率比较高时,会占用大量的CPU资源;

root@NanoPC-T6:/opt/qt-project/FloatVideo-TouchScreen# export DISPLAY=:0.0;./FloatVideo-TouchScreen -size 0.8
root@NanoPC-T6:~# top
任务: 278 total,   2 running, 276 sleeping,   0 stopped,   0 zombie
%Cpu(s): 36.0 us,  1.9 sy,  0.0 ni, 62.1 id,  0.0 wa,  0.0 hi,  0.0 si,  0.0 st
MiB Mem :  15953.1 total,  14749.5 free,    662.4 used,    541.2 buff/cache
MiB Swap:      0.0 total,      0.0 free,      0.0 used.  14995.4 avail Mem

 进程号 USER      PR  NI    VIRT    RES    SHR    %CPU  %MEM     TIME+ COMMAND
   1513 root      20   0 2876120 127804  72488 S 270.9   0.8   0:27.46 FloatVideo-Touc
    864 root      20   0 3345456 238548 186388 S  14.9   1.5   0:03.11 Xorg
   1251 pi        20   0 1861028  76424  57116 S   2.6   0.5   0:00.96 xfwm4
   ......

那么我们是不是可以通过GPU来实现图像的缩放呢,在RK3588上可以使用OpenCL接口进行GPU加速。

一、OpenCL环境搭建

OpenCL(Open Computing Language开放计算语言)是一种开放的、免版税的标准,用于超级计算机、云服务器、个人计算机、移动设备和嵌入式平台中各种加速器的跨平台并行编程。

OpenCL是由Khronos Group创建和管理的。OpenCL使应用程序能够使用系统或设备中的并行处理能力,从而使应用程序运行得更快、更流畅。

1.1 工作原理

OpenCL是一种编程框架和运行时,它使程序员能够创建称为内核程序(或内核)的小程序,这些程序可以在系统中的任何处理器上并行编译和执行。处理器可以是不同类型的任意组合,包括CPUGPUDSPFPGA或张量处理器,这就是为什么OpenCL经常被称为异构并行编程的解决方案。

OpenCL框架包含两个API

  • platform layer API:在主机CPU上运行,首先用于使程序能够发现系统中可用的并行处理器或计算设备。通过查询哪些计算设备可用,应用程序可以在不同的系统上便携地运行—适应加速器硬件的不同组合。一旦发现了计算设备,platform layer API就允许应用程序选择并初始化它想要使用的设备;
  • Runtime API:它使应用程序的内核程序能够为它们将要运行的计算设备编译,并行加载到这些处理器上并执行。一旦内核程序完成执行,将使用Runtime API收集结果;

为了更好适用于不同的处理器,OpenCL抽象出来了四大模型:

  • 平台模型:描述了OpenCL如何理解拓扑连接系统中的计算资源,对不同硬件及软件实现抽象,方便应用于不同设备;
  • 内存模型:对硬件的各种内存器进行了抽象;
  • 执行模型:程序是如何在硬件上执行的;
  • 编程模型:数据并行和任务并行;
1.2 平台模型

OpenCL中,需要一个主机处理器(Host),一般为CPU。而其它的硬件处理器(多核CPU/GPU/DSP等)被抽象成Compute Device

  • 每个Compute Device包含多个Compute Unit
  • 每个Compute Unit又包含多个Processing Elements(处理单元)。

举例说明:计算设备可以是GPU,计算单元对应于GPU内部的流多处理器(streaming multiprocessors(SMs)),处理单元对应于每个SM内部的单个流处理器。处理器通常通过共享指令调度和内存资源,以及增加本地处理器间通信,将处理单元分组为计算单元,以提高实现效率。

1.3 内存模型

OpenCL内存模型定义了如何访问和共享不同内核和处理单元之间的数据。

1.3.1 内存类型

OpenCL支持以下内存类型:

  • Global memory: 全局内存对在上下文中执行的所有工作项可访问,主机可以使用__global关键字读取、写入和映射命令访问全局内存,在单个工作组中,全局内存是一致的;
  • Constant memory:常量内存是用于主机分配和初始化的对象的内存区域, 所有工作项都可以以只读方式访问常量内存;
  • Local memory: 本地内存是特定于工作组的,工作组中的工作项可以访问本地内存;使用__local关键字进行访问,对于工作组中的所有工作项来说,本地内存是一致的;
  • Private memory:私有内存是特定于工作项的,其他工作项无法访问私有内存;
1.3.2 内存模型

OpenCL内存模型如下:

1.4 执行模型

OpenCL执行模型包括主机应用程序、上下文(context)和OpenCL内核的操作。

主机应用程序使用OpenCL命令队列将kernel和数据传输函数发送到设备以执行。

通过将命令入队到命令队列(Command Queues)中,kernel和数据传输函数可以与应用程序主机代码并行异步执行。

1.4.1 主机应用程序

主机应用程序在应用处理器上运行。主机应用程序通过为以下命令设置命令队列来管理内核的执行:

  • 内存命令;
  • 内核执行命令;
  • 同步操作;
1.4.2 上下文

主机应用程序为内核定义上下文。上下文包括:

  • 计算设备(Compute devices);

  • 内核(Kernels):OpenCL核心计算部分,类似C语言的代码。在需要设备执行计算任务时,数据会被推送到Compute Device,然后Compute Device的计算单元会并发执行内核程序;

  • 程序对象(Programs):Kernels的集合,OpenCL中可以使用cl_program表示;

  • 内存对象(Memory Objects.);

1.4.3 OpenCL内核的操作

Kernels在计算设备上运行。kernel是一段代码,在计算设备上与其它内核并行执行。内核的操作按以下顺序进行:

  • Kernels在主机应用程序中定义;
  • 主机应用程序将kernel提交给计算设备执行。计算设备可以是应用处理器、GPU或其它类型的处理器;
  • 当主机应用程序发出提交kernel的命令时,OpenCL创建工作项的NDRange
  • 对于NDRange中的每个元素,创建kernel的一个实例。这使得每个元素可以独立并行地进行处理。
1.5 OpenCL计算流程

对于OpenCl,利用显卡计算时,需要经历如下步骤:

  • 主机应用程序进行设备初始化(获取平台和设备id,创建上下文和命令队列);
  • 编写并编译kernel(读取内核文件->创建program对象->编译程序->创建内核) ;
  • 主机应用程序准备数据并传入设备(准备主机端数据,创建设备端内存对象并拷贝主机端数据);
  • 主机应用程序将kernel提交给设备执行(传入kernel函数参数, 启动kernel函数);
  • 将结果拷贝回主机应用程序;
  • 后续处理;
  • 释放资源。

二、OpenCL环境搭建

一个完整的OpenCL框架,从内核层到用户层,可分为四部分:

  • 内核层GPU驱动;
  • 用户层动态库;
  • 头文件;
  • 应用程序;
2.1 内核层GPU驱动

RK3588为例,搭载了Mail-G610 GPULinux内核提供了针对Mali-T6xx / Mali-T7xx / Mali-T8xx GPUGXX系列的Panfrost驱动,具体可以参考《Rockchip RK3399 - Mali-T860 GPU驱动》;

注意: 内核层GPU驱动这一部分,不需要自己移植,我们开发板所使用的的友善linux kernel 6.1已移植;

2.2 用户层动态库

用户层动态库有多种途径可以获得,比如以下两种:

  • 寻找官方(Mali ARM/Rockchip )提供的用户层动态库libmali.so

  • 下载KhronosGroup OpenCL-SDK源码,并编译,可以得到libOpenCL.so

下面我们分别介绍这两种方式, 对于这两种方式我们选择一种即可,对于我使用的NanoPC-T6采用第一种方式(默认已经支持)。

2.2.1 Mali ARM官方下载安装libmali.so

通过浏览器进入Mali ARM官网:https://developer.arm.com/downloads/-/mali-drivers/user-space

寻找官方提供的用户层动态库libmali.solibmali.so一般会有不同的版本(X11fbdevWayland等),其提供了opengleseglopencl接口。

不过不幸的是:Mail ARM官网并没有看到适用于RK3588的用户层动态库,但是RK3288的倒是有,这里我们就以RK3288为例:

下载后,解压缩可以看到:

注意:上图中libEGL.solibOpenCL.solibGLESv2.so等库大小均为0,不难猜测libmail.so应该提供了opengleseglopencl接口,也就是该库由以上几个库合并而成。

libmali.so存放在ARM/usr/lib/,同时建立软链接libOpenCL.so指向libmali.so

root@NanoPC-T6:~# ln -s /usr/lib/libmali.so /usr/lib/libOpenCL.so
2.2.2 Rockchip官方提供的libmali.so

我们使用的友善提供的debian文件系统已经安装了libmali.so,该用户层动态库是由Rockchip官方提供的。

如何来查看是否已经安装了OpenCL库和驱动,可以通过如下命令检查是否已经安装了libmali.so

root@NanoPC-T6:~# find /usr -name libmali.so
/usr/lib/aarch64-linux-gnu/libmali.so
root@NanoPC-T6:~# strings /usr/lib/aarch64-linux-gnu/libmali.so | grep Mali-G610
Mali-G610

root@NanoPC-T6:~# strings /usr/lib/aarch64-linux-gnu/libmali.so | grep cl
.....
clReleaseCommandBufferKHR
clReleaseCommandQueue
clReleaseContext
clReleaseDevice
clReleaseEvent
clReleaseKernel
clReleaseMemObject
.....

root@NanoPC-T6:~# ls -l /usr/lib/aarch64-linux-gnu/libmali.so
lrwxrwxrwx 1 root root 12  7月 29  2020 /usr/lib/aarch64-linux-gnu/libmali.so -> libmali.so.1

其中/usr/lib/aarch64-linux-gnu/libmali.solibmali.so库的路径,Mali-G610Mali GPU驱动的版本号。

如果命令输出为空,则说明该库不是Mali GPU驱动库。如果输出包含Mali-G610 字符串,则说明该库是Mali GPU驱动库,并且版本号为Mali-G610

此外在/usr/lib/aarch64-linux-gnu目录下包含单独的opengleseglopencl库;

root@NanoPC-T6:/opt# ls -l /usr/lib/aarch64-linux-gnu/libOpenCL*
lrwxrwxrwx 1 root root    18  1月 12  2021 /usr/lib/aarch64-linux-gnu/libOpenCL.so.1 -> libOpenCL.so.1.0.0
-rw-r--r-- 1 root root 60856  1月 12  2021 /usr/lib/aarch64-linux-gnu/libOpenCL.so.1.0.0

root@NanoPC-T6:/opt# strings /usr/lib/aarch64-linux-gnu/libOpenCL.so.1.0.0 | grep cl
fclose
closedir
dlclose
clGetExtensionFunctionAddress
clGetPlatformIDs
clCreateContext
clCreateContextFromType
clGetGLContextInfoKHR
......

root@NanoPC-T6:/opt# ls -l /usr/lib/aarch64-linux-gnu/libEGL*
lrwxrwxrwx 1 root root     20  3月 25  2021 /usr/lib/aarch64-linux-gnu/libEGL_mesa.so.0 -> libEGL_mesa.so.0.0.0
-rw-r--r-- 1 root root 259072  3月 25  2021 /usr/lib/aarch64-linux-gnu/libEGL_mesa.so.0.0.0
lrwxrwxrwx 1 root root     11  7月 29  2020 /usr/lib/aarch64-linux-gnu/libEGL.so -> libEGL.so.1
lrwxrwxrwx 1 root root     15  7月 29  2020 /usr/lib/aarch64-linux-gnu/libEGL.so.1 -> libEGL.so.1.1.0
-rw-r--r-- 1 root root  84416  7月 29  2020 /usr/lib/aarch64-linux-gnu/libEGL.so.1.1.0
......

也可以通过如下clinfo命令查看是否已经安装OpenCL库,如果出现下图所示界面,则系统已经安装;

root@NanoPC-T6:~# aptitude install clinfo
root@NanoPC-T6:~# clinfo
arm_release_ver: g13p0-01eac0, rk_so_ver: 10
Number of platforms                               1
  Platform Name                                   ARM Platform
  Platform Vendor                                 ARM
  Platform Version                                OpenCL 3.0 v1.g13p0-01eac0.a8b6f0c7e1f83c654c60d1775112dbe4
  Platform Profile                                FULL_PROFILE
  Platform Extensions                             cl_khr_global_int32_base_atomics cl_khr_global_int32_extended_atomics           ......
NULL platform behavior
  clGetPlatformInfo(NULL, CL_PLATFORM_NAME, ...)  ARM Platform
  clGetDeviceIDs(NULL, CL_DEVICE_TYPE_ALL, ...)   Success [ARM]
  clCreateContext(NULL, ...) [default]            Success [ARM]
  clCreateContextFromType(NULL, CL_DEVICE_TYPE_DEFAULT)  Success (1)
    Platform Name                                 ARM Platform
    Device Name                                   Mali-G610 r0p0            # GPU型号
  clCreateContextFromType(NULL, CL_DEVICE_TYPE_CPU)  No devices found in platform
  clCreateContextFromType(NULL, CL_DEVICE_TYPE_GPU)  Success (1)
    Platform Name                                 ARM Platform
    Device Name                                   Mali-G610 r0p0
  clCreateContextFromType(NULL, CL_DEVICE_TYPE_ACCELERATOR)  No devices found in platform
  clCreateContextFromType(NULL, CL_DEVICE_TYPE_CUSTOM)  No devices found in platform
  clCreateContextFromType(NULL, CL_DEVICE_TYPE_ALL)  Success (1)
    Platform Name                                 ARM Platform
    Device Name                                   Mali-G610 r0p0

ICD loader properties
  ICD loader Name                                 OpenCL ICD Loader
  ICD loader Vendor                               OCL Icd free software
  ICD loader Version                              2.2.14
  ICD loader Profile                              OpenCL 3.0

接着我们需要将建立软链接libOpenCL.so指向libmali.so

root@NanoPC-T6:~# ln -s /usr/lib/aarch64-linux-gnu/libmali.so /usr/lib/aarch64-linux-gnu/libOpenCL.so
root@NanoPC-T6:~# ls -l /usr/lib/aarch64-linux-gnu/libOpenCL.so
lrwxrwxrwx 1 root root 37  1月 16 23:43 /usr/lib/aarch64-linux-gnu/libOpenCL.so -> /usr/lib/aarch64-linux-gnu/libmali.so
2.2.3 OpenCL SDK编译安装

如果没有安装,请按照如下步骤安装:下载OpenCL SDK进行编译安装,具体可以参考《OpenCL安装过程记录》。

Khronos GroupOpenCL SDK是一个通用的官方开发工具包,适用于多个硬件平台,而AMDIntel等硬件供应商提供的OpenCL SDK则更专注于其特定硬件平台的优化和支持。根据您的需求和使用的硬件平台,选择适合的OpenCL SDK可以帮助您获得最佳的性能和开发体验。

下载源码:

root@NanoPC-T6:/opt# git clone --recursive https://github.com/KhronosGroup/OpenCL-SDK.git

运行以下命令来配置构建过程,并指定安装路径为/opt/OpenCL

root@NanoPC-T6:/opt/OpenCL-SDK# cmake -S . -B build -DCMAKE_INSTALL_PREFIX=/opt/OpenCL
-- The C compiler identification is GNU 10.2.1
-- The CXX compiler identification is GNU 10.2.1
-- Detecting C compiler ABI info
-- Detecting C compiler ABI info - done
-- Check for working C compiler: /usr/bin/cc - skipped
-- Detecting C compile features
-- Detecting C compile features - done
-- Detecting CXX compiler ABI info
-- Detecting CXX compiler ABI info - done
-- Check for working CXX compiler: /usr/bin/c++ - skipped
-- Detecting CXX compile features
-- Detecting CXX compile features - done
-- No build type selected, default to Release
-- Looking for pthread.h
-- Looking for pthread.h - found
-- Performing Test CMAKE_HAVE_LIBC_PTHREAD
-- Performing Test CMAKE_HAVE_LIBC_PTHREAD - Failed
-- Looking for pthread_create in pthreads
-- Looking for pthread_create in pthreads - not found
-- Looking for pthread_create in pthread
-- Looking for pthread_create in pthread - found
-- Found Threads: TRUE
-- Looking for secure_getenv
-- Looking for secure_getenv - found
-- Looking for __secure_getenv
-- Looking for __secure_getenv - not found
-- Check if compiler accepts -pthread
-- Check if compiler accepts -pthread - yes
-- Could NOT find Doxygen (missing: DOXYGEN_EXECUTABLE)
-- cargs (https://521github.com/likle/cargs) not found. To self-host, set cargs_INCLUDE_PATH and cargs_LIBRARY to point to the headers and library respectively adding '-D cargs_INCLUDE_PATH=/path/to/cargs/include/dir -D cargs_LIBRARY/path/to/cargs/libcargs' to the cmake command. (missing: cargs_INCLUDE_PATH cargs_LIBRARY)
-- Fetching cargs.
-- Adding cargs subproject: /opt/OpenCL-SDK/build/_deps/cargs-external-src
-- TCLAP (http://tclap.sourceforge.net/) not found. To self-host, set TCLAP_INCLUDE_PATH to point to the headers adding '-DTCLAP_INCLUDE_PATH=/path/to/tclap' to the cmake command. (missing: TCLAP_INCLUDE_PATH)
-- Fetching TCLAP.
-- Found TCLAP: /opt/OpenCL-SDK/build/_deps/tclap-external-src/include
-- Stb (https://521github.com/nothings/stb) not found. To self-host, set Stb_INCLUDE_PATH to point to the headers adding '-D Stb_INCLUDE_PATH=/path/to/stb' to the cmake command. (missing: Stb_INCLUDE_PATH)
-- Fetching Stb.
-- Found Stb: /opt/OpenCL-SDK/build/_deps/stb-external-src
-- Found X11: /usr/include
-- Looking for XOpenDisplay in /usr/lib/aarch64-linux-gnu/libX11.so;/usr/lib/aarch64-linux-gnu/libXext.so
-- Looking for XOpenDisplay in /usr/lib/aarch64-linux-gnu/libX11.so;/usr/lib/aarch64-linux-gnu/libXext.so - found
-- Looking for gethostbyname
-- Looking for gethostbyname - found
-- Looking for connect
-- Looking for connect - found
-- Looking for remove
-- Looking for remove - found
-- Looking for shmat
-- Looking for shmat - found
-- Looking for IceConnectionNumber in ICE
-- Looking for IceConnectionNumber in ICE - found
-- Could NOT find glm (missing: glm_DIR)
-- Fetching glm.
-- Adding glm subproject: /opt/OpenCL-SDK/build/_deps/glm-external-src
CMake Warning (dev) at /usr/share/cmake-3.18/Modules/FindOpenGL.cmake:305 (message):
  Policy CMP0072 is not set: FindOpenGL prefers GLVND by default when
  available.  Run "cmake --help-policy CMP0072" for policy details.  Use the
  cmake_policy command to set the policy and suppress this warning.

  FindOpenGL found both a legacy GL library:

    OPENGL_gl_LIBRARY: /usr/lib/aarch64-linux-gnu/libGL.so

  and GLVND libraries for OpenGL and GLX:

    OPENGL_opengl_LIBRARY: /usr/lib/aarch64-linux-gnu/libOpenGL.so
    OPENGL_glx_LIBRARY: /usr/lib/aarch64-linux-gnu/libGLX.so

  OpenGL_GL_PREFERENCE has not been set to "GLVND" or "LEGACY", so for
  compatibility with CMake 3.10 and below the legacy GL library will be used.
Call Stack (most recent call first):
  cmake/Dependencies/OpenGL/OpenGL.cmake:1 (find_package)
  cmake/Dependencies.cmake:17 (include)
  CMakeLists.txt:50 (include)
This warning is for project developers.  Use -Wno-dev to suppress it.

-- Found OpenGL: /usr/lib/aarch64-linux-gnu/libOpenGL.so
-- Could NOT find GLEW (missing: GLEW_INCLUDE_DIRS GLEW_LIBRARIES)
-- Fetching GLEW.
-- Adding GLEW subproject: /opt/OpenCL-SDK/build/_deps/glew-external-src
CMake Warning (dev) at build/_deps/glew-external-src/CMakeLists.txt:2 (project):
  Policy CMP0048 is not set: project() command manages VERSION variables.
  Run "cmake --help-policy CMP0048" for policy details.  Use the cmake_policy
  command to set the policy and suppress this warning.

  The following variable(s) would be set to empty:

    PROJECT_VERSION
    PROJECT_VERSION_MAJOR
    PROJECT_VERSION_MINOR
    PROJECT_VERSION_PATCH
This warning is for project developers.  Use -Wno-dev to suppress it.

-- Found Freetype: /usr/lib/aarch64-linux-gnu/libfreetype.so (found version "2.10.4")
-- Fetching SFML.
-- Adding SFML subproject: /opt/OpenCL-SDK/build/_deps/sfml-external-src
-- libudev stable: 1
-- Found UDev: /usr/lib/aarch64-linux-gnu/libudev.so
--    include: /usr/include
-- Performing Test COMPILER_HAS_HIDDEN_VISIBILITY
-- Performing Test COMPILER_HAS_HIDDEN_VISIBILITY - Success
-- Performing Test COMPILER_HAS_HIDDEN_INLINE_VISIBILITY
-- Performing Test COMPILER_HAS_HIDDEN_INLINE_VISIBILITY - Success
-- Performing Test COMPILER_HAS_DEPRECATED_ATTR
-- Performing Test COMPILER_HAS_DEPRECATED_ATTR - Success
-- Looking for sin in m
-- Looking for sin in m - found
-- Configuring done
-- Generating done
-- Build files have been written to: /opt/OpenCL-SDK/build
root@NanoPC-T6:/opt/OpenCL-SDK#

其中:

  • -S .:指定源代码目录的路径;
  • -B build:指定构建目录的路径;
  • -DCMAKE_INSTALL_PREFIX=/opt/OpenCL:指定cmake执行install 目标时,安装的路径前缀;

接着运行以下命令在./build目录下执行构建操作,只构建install目标,将生成的文件安装到指定的位置;

root@NanoPC-T6:/opt/OpenCL-SDK# cmake --build build --target install

编译完成之后,我们查看安装目录:

root@NanoPC-T6:/opt/OpenCL-SDK# ls /opt/OpenCL -l
总用量 16
drwxr-xr-x 2 root root 4096  1月 16 20:44 bin  
drwxr-xr-x 5 root root 4096  1月 16 20:44 include # 头文件
drwxr-xr-x 4 root root 4096  1月 16 20:44 lib   # 库文件
drwxr-xr-x 5 root root 4096  1月 16 20:44 share

root@NanoPC-T6:/opt/OpenCL-SDK# ls -l /opt/OpenCL/lib
总用量 4564
drwxr-xr-x 4 root root    4096  1月 16 20:44 cmake
-rw-r--r-- 1 root root    4842  1月 16 20:41 libcargs.a
-rw-r--r-- 1 root root 1215506  1月 16 20:42 libglew.a
lrwxrwxrwx 1 root root      23  1月 16 20:44 libglew-shared.so -> libglew-shared.so.2.2.0
-rw-r--r-- 1 root root  961392  1月 16 20:42 libglew-shared.so.2.2.0
-rw-r--r-- 1 root root   90550  1月 16 20:43 libOpenCLExt.a
-rw-r--r-- 1 root root 1269528  1月 16 20:43 libOpenCLSDKCpp.so
-rw-r--r-- 1 root root  205392  1月 16 20:43 libOpenCLSDK.so
lrwxrwxrwx 1 root root      14  1月 16 20:44 libOpenCL.so -> libOpenCL.so.1
lrwxrwxrwx 1 root root      16  1月 16 20:44 libOpenCL.so.1 -> libOpenCL.so.1.2
-rw-r--r-- 1 root root   74744  1月 16 20:41 libOpenCL.so.1.2
-rw-r--r-- 1 root root   61152  1月 16 20:42 libOpenCLUtilsCpp.so
-rw-r--r-- 1 root root   27096  1月 16 20:42 libOpenCLUtils.so
lrwxrwxrwx 1 root root      23  1月 16 20:44 libsfml-graphics.so -> libsfml-graphics.so.2.5
lrwxrwxrwx 1 root root      25  1月 16 20:44 libsfml-graphics.so.2.5 -> libsfml-graphics.so.2.5.1
-rw-r--r-- 1 root root  456128  1月 16 20:42 libsfml-graphics.so.2.5.1
lrwxrwxrwx 1 root root      21  1月 16 20:44 libsfml-system.so -> libsfml-system.so.2.5
lrwxrwxrwx 1 root root      23  1月 16 20:44 libsfml-system.so.2.5 -> libsfml-system.so.2.5.1
-rw-r--r-- 1 root root   71592  1月 16 20:42 libsfml-system.so.2.5.1
lrwxrwxrwx 1 root root      21  1月 16 20:44 libsfml-window.so -> libsfml-window.so.2.5
lrwxrwxrwx 1 root root      23  1月 16 20:44 libsfml-window.so.2.5 -> libsfml-window.so.2.5.1
-rw-r--r-- 1 root root  202536  1月 16 20:42 libsfml-window.so.2.5.1
drwxr-xr-x 2 root root    4096  1月 16 20:44 pkgconfig
root@NanoPC-T6:/opt/OpenCL-SDK# ls -l /opt/OpenCL/include/
总用量 20
-rw-r--r-- 1 root root 4553  1月 16 20:40 cargs.h
drwxr-xr-x 3 root root 4096  1月 16 20:44 CL
drwxr-xr-x 2 root root 4096  1月 16 20:44 GL
drwxr-xr-x 7 root root 4096  1月 16 20:44 SFML
root@NanoPC-T6:/opt/OpenCL-SDK# ls -l /opt/OpenCL/include/CL/
总用量 788
-rw-r--r-- 1 root root    786  1月 16 20:38 cl2.hpp
-rw-r--r-- 1 root root   8057  1月 16 20:38 cl_d3d10.h
-rw-r--r-- 1 root root   8095  1月 16 20:38 cl_d3d11.h
-rw-r--r-- 1 root root  12246  1月 16 20:38 cl_dx9_media_sharing.h
-rw-r--r-- 1 root root    959  1月 16 20:38 cl_dx9_media_sharing_intel.h
-rw-r--r-- 1 root root   5672  1月 16 20:38 cl_egl.h
-rw-r--r-- 1 root root 127490  1月 16 20:38 cl_ext.h
-rw-r--r-- 1 root root    902  1月 16 20:38 cl_ext_intel.h
-rw-r--r-- 1 root root  33387  1月 16 20:38 cl_function_types.h
-rw-r--r-- 1 root root    905  1月 16 20:38 cl_gl_ext.h
-rw-r--r-- 1 root root  12040  1月 16 20:38 cl_gl.h
-rw-r--r-- 1 root root  81631  1月 16 20:38 cl.h
-rw-r--r-- 1 root root  10430  1月 16 20:38 cl_half.h
-rw-r--r-- 1 root root  11505  1月 16 20:38 cl_icd.h
-rw-r--r-- 1 root root   3544  1月 16 20:38 cl_layer.h
-rw-r--r-- 1 root root  43430  1月 16 20:38 cl_platform.h
-rw-r--r-- 1 root root   7090  1月 16 20:38 cl_va_api_media_sharing_intel.h
-rw-r--r-- 1 root root   3125  1月 16 20:38 cl_version.h
-rw-r--r-- 1 root root    970  1月 16 20:38 opencl.h
-rw-r--r-- 1 root root 396735  1月 16 20:38 opencl.hpp
drwxr-xr-x 2 root root   4096  1月 16 20:44 Utils

接着我们将库文件和头文件放置到/usr路径下:

sudo ln -s /opt/OpenCL/include/CL /usr/include
sudo ln -s /opt/OpenCL/include/GL /usr/include
sudo ln -s /opt/OpenCL/include/SFML /usr/include
sudo ln -s /opt/OptnCL/lib/libOpenCL.so /usr/lib
2.3 安装头文件

从官网下载头文件OpenCL-Headers

root@NanoPC-T6:/opt# git clone https://github.com/extdomains/github.com/KhronosGroup/OpenCL-Headers.git

运行以下命令来配置构建过程,并指定安装路径为/usr

root@NanoPC-T6:/opt/OpenCL-Headers# cmake -S . -B build -DCMAKE_INSTALL_PREFIX=/usr
-- The C compiler identification is GNU 10.2.1
-- Detecting C compiler ABI info
-- Detecting C compiler ABI info - done
-- Check for working C compiler: /usr/bin/cc - skipped
-- Detecting C compile features
-- Detecting C compile features - done
-- The CXX compiler identification is GNU 10.2.1
-- Detecting CXX compiler ABI info
-- Detecting CXX compiler ABI info - done
-- Check for working CXX compiler: /usr/bin/c++ - skipped
-- Detecting CXX compile features
-- Detecting CXX compile features - done
-- Found Python3: /usr/bin/python3.9 (found version "3.9.2") found components: Interpreter
-- Configuring done
-- Generating done
-- Build files have been written to: /opt/OpenCL-Headers/build

其中:

  • -S .:指定源代码目录的路径;
  • -B build:指定构建目录的路径;
  • -DCMAKE_INSTALL_PREFIX=/usr:指定cmake执行install 目标时,安装的路径前缀;

如上命令会让cmake.目录下查找CMakeLists.txt文件,并在./build目录下生成Makefile文件。

接着运行以下命令在./build目录下执行构建操作,只构建install目标,将生成的文件安装到指定的位置;

root@NanoPC-T6:/opt/OpenCL-Headers# cmake --build build --target install
Scanning dependencies of target headers_c_200
[  0%] Building C object tests/lang_c/CMakeFiles/headers_c_200.dir/__/test_headers.c.o
[  0%] Linking C executable headers_c_200
[  0%] Built target headers_c_200
Scanning dependencies of target headers_c_120
[  1%] Building C object tests/lang_c/CMakeFiles/headers_c_120.dir/__/test_headers.c.o
[  1%] Linking C executable headers_c_120
[  1%] Built target headers_c_120
Scanning dependencies of target cl_version_h_c_300
[  1%] Building C object tests/lang_c/CMakeFiles/cl_version_h_c_300.dir/__/test_cl_version.h.c.o
[  2%] Linking C executable cl_version_h_c_300
.......
[ 99%] Built target cl_egl_h_cpp_100
Scanning dependencies of target cl_gl_h_cpp_120
[100%] Building CXX object tests/lang_cpp/CMakeFiles/cl_gl_h_cpp_120.dir/test_cl_gl.h.cpp.o
[100%] Linking CXX executable cl_gl_h_cpp_120
[100%] Built target cl_gl_h_cpp_120
Install the project...
-- Install configuration: ""
-- Installing: /usr/include/CL
-- Installing: /usr/include/CL/opencl.h
-- Installing: /usr/include/CL/cl_egl.h
-- Installing: /usr/include/CL/cl_ext_intel.h
-- Installing: /usr/include/CL/cl_layer.h
-- Installing: /usr/include/CL/cl_platform.h
-- Installing: /usr/include/CL/cl_d3d10.h
-- Installing: /usr/include/CL/cl_va_api_media_sharing_intel.h
-- Installing: /usr/include/CL/cl_icd.h
-- Installing: /usr/include/CL/cl.h
-- Installing: /usr/include/CL/cl_function_types.h
-- Installing: /usr/include/CL/cl_dx9_media_sharing.h
-- Installing: /usr/include/CL/cl_dx9_media_sharing_intel.h
-- Installing: /usr/include/CL/cl_gl_ext.h
-- Installing: /usr/include/CL/cl_d3d11.h
-- Installing: /usr/include/CL/cl_version.h
-- Installing: /usr/include/CL/cl_half.h
-- Installing: /usr/include/CL/cl_ext.h
-- Installing: /usr/include/CL/cl_gl.h
-- Installing: /usr/share/cmake/OpenCLHeaders/OpenCLHeadersTargets.cmake
-- Installing: /usr/share/cmake/OpenCLHeaders/OpenCLHeadersConfig.cmake
-- Installing: /usr/share/cmake/OpenCLHeaders/OpenCLHeadersConfigVersion.cmake
-- Installing: /usr/share/pkgconfig/OpenCL-Headers.pc

头文件已经安装到/usr/include/CL目录下:

root@NanoPC-T6:/opt/OpenCL-Headers# ls -l /usr/include/CL
总用量 392
-rw-r--r-- 1 root root   8057  1月 15 00:10 cl_d3d10.h
-rw-r--r-- 1 root root   8095  1月 15 00:10 cl_d3d11.h
-rw-r--r-- 1 root root  12246  1月 15 00:10 cl_dx9_media_sharing.h
-rw-r--r-- 1 root root    959  1月 15 00:10 cl_dx9_media_sharing_intel.h
-rw-r--r-- 1 root root   5672  1月 15 00:10 cl_egl.h
-rw-r--r-- 1 root root 127490  1月 15 00:10 cl_ext.h
-rw-r--r-- 1 root root    902  1月 15 00:10 cl_ext_intel.h
-rw-r--r-- 1 root root  33387  1月 15 00:10 cl_function_types.h
-rw-r--r-- 1 root root    905  1月 15 00:10 cl_gl_ext.h
-rw-r--r-- 1 root root  12040  1月 15 00:10 cl_gl.h
-rw-r--r-- 1 root root  81631  1月 15 00:10 cl.h
-rw-r--r-- 1 root root  10430  1月 15 00:10 cl_half.h
-rw-r--r-- 1 root root  11505  1月 15 00:10 cl_icd.h
-rw-r--r-- 1 root root   3544  1月 15 00:10 cl_layer.h
-rw-r--r-- 1 root root  43430  1月 15 00:10 cl_platform.h
-rw-r--r-- 1 root root   7090  1月 15 00:10 cl_va_api_media_sharing_intel.h
-rw-r--r-- 1 root root   3125  1月 15 00:10 cl_version.h
-rw-r--r-- 1 root root    970  1月 15 00:10 opencl.h

三、OpenCL测试

此时已经有动态库和头文件,可以进行测试了。在/opt/目录下创建opencl-project文件夹;

root@NanoPC-T6:/opt# mkdir opencl-project

接着创建platform文件夹;

root@NanoPC-T6:/opt# cd opencl-project/
root@NanoPC-T6:/opt/opencl-project# mkdir platform
root@NanoPC-T6:/opt/opencl-project# cd platform
3.1 platform.cpp

/opt/opencl-project/platform目录下编写测试代码platform.cpp

#include <stdio.h>
#include <stdlib.h>
#include <CL/cl.h>

#define MAX_PLATFORMS 10
#define MAX_DEVICES 10

int main() {
    cl_platform_id platforms[MAX_PLATFORMS];
    cl_device_id devices[MAX_DEVICES];
    cl_uint num_platforms, num_devices;
    cl_context context;
    cl_command_queue command_queue;
    cl_program program;
    cl_kernel kernel;
    cl_int ret;

    // 获取平台数量
    ret = clGetPlatformIDs(MAX_PLATFORMS, platforms, &num_platforms);
    if (ret != CL_SUCCESS) {
        printf("Failed to get platform IDs
");
        return -1;
    }

    printf("Number of platforms: %u
", num_platforms);

    // 遍历打印平台信息
    for (cl_uint i = 0; i < num_platforms; i++) {
        char platform_name[128];
        char platform_vendor[128];

        ret = clGetPlatformInfo(platforms[i], CL_PLATFORM_NAME, sizeof(platform_name), platform_name, NULL);
        if (ret != CL_SUCCESS) {
            printf("Failed to get platform name for platform %u
", i);
        }

        ret = clGetPlatformInfo(platforms[i], CL_PLATFORM_VENDOR, sizeof(platform_vendor), platform_vendor, NULL);
        if (ret != CL_SUCCESS) {
            printf("Failed to get platform vendor for platform %u
", i);
        }

        printf("Platform %u:
", i);
        printf("    Name: %s
", platform_name);
        printf("    Vendor: %s
", platform_vendor);
        printf("
");
    }

    // 获取设备数量
    ret = clGetDeviceIDs(platforms[0], CL_DEVICE_TYPE_GPU, MAX_DEVICES, devices, &num_devices);
    if (ret != CL_SUCCESS) {
        printf("Failed to get device IDs
");
        return -1;
    }

    // 创建OpenCL上下文
    context = clCreateContext(NULL, num_devices, devices, NULL, NULL, &ret);
    if (ret != CL_SUCCESS) {
        printf("Failed to create context
");
        return -1;
    }

    // 创建命令队列
    command_queue = clCreateCommandQueue(context, devices[0], 0, &ret);
    if (ret != CL_SUCCESS) {
        printf("Failed to create command queue
");
        return -1;
    }

    // 定义和构建OpenCL内核
    const char *kernel_source = "__kernel void hello_world() {
"
                                "    printf("Hello, World!\n");
"
                                "}
";
    program = clCreateProgramWithSource(context, 1, &kernel_source, NULL, &ret);
    if (ret != CL_SUCCESS) {
        printf("Failed to create program
");
        return -1;
    }

    ret = clBuildProgram(program, num_devices, devices, NULL, NULL, NULL);
    if (ret != CL_SUCCESS) {
        printf("Failed to build program
");
        return -1;
    }

    // 创建OpenCL内核对象
    kernel = clCreateKernel(program, "hello_world", &ret);
    if (ret != CL_SUCCESS) {
        printf("Failed to create kernel
");
        return -1;
    }

    // 执行内核函数
    ret = clEnqueueTask(command_queue, kernel, 0, NULL, NULL);
    if (ret != CL_SUCCESS) {
        printf("Failed to enqueue task
");
        return -1;
    }

    // 等待执行完成
    ret = clFinish(command_queue);
    if (ret != CL_SUCCESS) {
        printf("Failed to finish execution
");
        return -1;
    }

    printf("Kernel executed successfully
");

    // 清理资源
    ret = clReleaseKernel(kernel);
    ret = clReleaseProgram(program);
    ret = clReleaseCommandQueue(command_queue);
    ret = clReleaseContext(context);

    return 0;
}
3.2 编译

这里我们介绍两种源码编译的方式。

3.2.1 直接编译

我们可以直接执行如下编译命令:

root@NanoPC-T6:/opt/opencl-project/platform# gcc platform.cpp -o platform -lmali

-lmail用于链接libmali.so库文件,-l选项指定要链接的库文件名,并在文件名前加上lib.so的前缀和后缀。所以-lmali告诉编译器要链接的库文件名为libmali.so

那么编译器如何知道libmali.so在哪里的呢?

  • 首先搜索预定义的默认路径,如/usr/lib/usr/local/lib等;
  • 如果共享库没有在这些路径中找到,则会搜索在/etc/ld.so.conf/etc/ld.so.conf.d目录中指定的路径。这些路径可以包含自定义共享库路径,比如:
root@NanoPC-T6:/opt/opencl-project/platform# ls -l /etc/ld.so.conf.d/
总用量 12
-rw-r--r-- 1 root root  32  7月 29  2020 00-aarch64-mali.conf
-rw-r--r-- 1 root root 103  4月 20  2023 aarch64-linux-gnu.conf
-rw-r--r-- 1 root root  44  9月 23  2022 libc.conf
root@NanoPC-T6:/opt/opencl-project/platform#  cat /etc/ld.so.conf.d/aarch64-linux-gnu.conf
# Multiarch support
/usr/local/lib/aarch64-linux-gnu
/lib/aarch64-linux-gnu
/usr/lib/aarch64-linux-gnu     # 该路径下有libmali.so库文件
3.2.2 cmake编译

当然也可以使用cmake进行编译platform.cpp,接下来我们介绍cmake编译配置。

(1) 在/opt/opencl-project/platform目录下创建CMakeLists.txt

cmake_minimum_required(VERSION 3.0)
cmake_policy(VERSION 3.0...3.18.4)
project(proj)
add_executable(platform platform.cpp)
#寻找OpenCL库  /usr/share/cmake-3.18/Modules/FindOpenCL.cmake
find_package(OpenCL REQUIRED)
#打印调试信息
MESSAGE(STATUS "Project: ${PROJECT_NAME}")
MESSAGE(STATUS "OpenCL library status:")
MESSAGE(STATUS "    version: ${OpenCL_VERSION_STRING}")
MESSAGE(STATUS "    libraries: ${OpenCL_LIBRARY}")
MESSAGE(STATUS "    include path: ${OpenCL_INCLUDE_DIR}")

target_link_libraries(platform PRIVATE OpenCL::OpenCL)

(2) 配置构建过程:

root@NanoPC-T6:/opt/opencl-project/platform#  cmake -S . -B build
-- The C compiler identification is GNU 10.2.1
-- The CXX compiler identification is GNU 10.2.1
-- Detecting C compiler ABI info
-- Detecting C compiler ABI info - done
-- Check for working C compiler: /usr/bin/cc - skipped
-- Detecting C compile features
-- Detecting C compile features - done
-- Detecting CXX compiler ABI info
-- Detecting CXX compiler ABI info - done
-- Check for working CXX compiler: /usr/bin/c++ - skipped
-- Detecting CXX compile features
-- Detecting CXX compile features - done
-- Looking for CL_VERSION_2_2
-- Looking for CL_VERSION_2_2 - found
-- Found OpenCL: /usr/lib/aarch64-linux-gnu/libOpenCL.so (found version "2.2")
-- Project: proj
-- OpenCL library status:
--     version: 2.2 
--     libraries: /usr/lib/aarch64-linux-gnu/libOpenCL.so         # 库文件路径
--     include path: /usr/include  # 头文件路径
-- Configuring done
-- Generating done
-- Build files have been written to: /opt/opencl-project/platform/build

其中:

  • -S .:选项指定源代码目录的路径,CMake将在该路径下查找CMakeLists.txt文件;
  • -B build:选项指定构建目录的路径;

实际上我们使用的版本是OpenCL 3.0,这里判定为2.2版本是因为cmake version 3.18.4 FindOpenCL.cmake能够识别的最大版本为2.2,其通过在CL/cl.h文件查找CL_VERSION_${VERSION}宏来判定安装的版本的。

可以通过修改/usr/share/cmake-3.18/Modules/FindOpenCL.cmake解决这个问题:

foreach(VERSION "3_0" "2_2" "2_1" "2_0" "1_2" "1_1" "1_0")

(3) 执行构建操作,生成可执行程序platform

root@NanoPC-T6:/opt/OpenCL-Headers/exmaples# cmake --build build
Scanning dependencies of target platform
[ 50%] Building CXX object CMakeFiles/platform.dir/platform.cpp.o
In file included from /usr/include/CL/cl.h:20,
                 from /usr/include/CL/opencl.h:24,
                 from /opt/OpenCL-Headers/exmaples/platform.cpp:1:
/usr/include/CL/cl_version.h:22:104: note: ‘#pragma message: cl_version.h: CL_TARGET_OPENCL_VERSION is not defined. Defaulting to 300 (OpenCL 3.0)’
   22 | #pragma message("cl_version.h: CL_TARGET_OPENCL_VERSION is not defined. Defaulting to 300 (OpenCL 3.0)")
      |                                                                                                        ^
[100%] Linking CXX executable platform
[100%] Built target platform

执行程序:

root@NanoPC-T6:/opt/opencl-project/platform# ls -l build/
总用量 48
-rw-r--r-- 1 root root 14229  1月 16 23:45 CMakeCache.txt
drwxr-xr-x 5 root root  4096  1月 16 23:46 CMakeFiles
-rw-r--r-- 1 root root  1632  1月 16 23:45 cmake_install.cmake
-rw-r--r-- 1 root root  5253  1月 16 23:45 Makefile
-rwxr-xr-x 1 root root 14248  1月 16 23:46 platform

root@NanoPC-T6:/opt/opencl-project/platform# ./build/platform
arm_release_ver: g13p0-01eac0, rk_so_ver: 10
Number of platforms: 1
Platform 0:
    Name: ARM Platform
    Vendor: ARM

Kernel executed successfully

四、OpenCV测试用例

/opt/opencl-project目录下新建opencv-ocl项目,源码位于:https://521github.com/opencv/opencv/tree/3.4.0/samples/opencl

4.1 OCL介绍
4.2 项目源码
4.2.1 main.c
点击查看代码
/*
// The example of interoperability between OpenCL and OpenCV.
// This will loop through frames of video either from input media file
// or camera device and do processing of these data in OpenCL and then
// in OpenCV. In OpenCL it does inversion of pixels in left half of frame and
// in OpenCV it does bluring in the right half of frame.
*/
#include <cstdio>
#include <cstdlib>
#include <iostream>
#include <fstream>
#include <string>
#include <sstream>
#include <iomanip>
#include <stdexcept>

#define CL_USE_DEPRECATED_OPENCL_2_0_APIS // eliminate build warning

#if __APPLE__
#include <OpenCL/cl.h>
#else
#include <CL/cl.h>
#endif

#include <opencv2/core/ocl.hpp>
#include <opencv2/core/utility.hpp>
#include <opencv2/video.hpp>
#include <opencv2/highgui.hpp>
#include <opencv2/imgproc.hpp>


using namespace std;
using namespace cv;

namespace opencl {

class PlatformInfo
{
public:
    PlatformInfo()
    {}

    ~PlatformInfo()
    {}

    cl_int QueryInfo(cl_platform_id id)
    {
        query_param(id, CL_PLATFORM_PROFILE, m_profile);
        query_param(id, CL_PLATFORM_VERSION, m_version);
        query_param(id, CL_PLATFORM_NAME, m_name);
        query_param(id, CL_PLATFORM_VENDOR, m_vendor);
        query_param(id, CL_PLATFORM_EXTENSIONS, m_extensions);
        return CL_SUCCESS;
    }

    std::string Profile()    { return m_profile; }
    std::string Version()    { return m_version; }
    std::string Name()       { return m_name; }
    std::string Vendor()     { return m_vendor; }
    std::string Extensions() { return m_extensions; }

private:
    cl_int query_param(cl_platform_id id, cl_platform_info param, std::string& paramStr)
    {
        cl_int res;

        size_t psize;
        cv::AutoBuffer<char> buf;

        res = clGetPlatformInfo(id, param, 0, 0, &psize);
        if (CL_SUCCESS != res)
            throw std::runtime_error(std::string("clGetPlatformInfo failed"));

        buf.resize(psize);
        res = clGetPlatformInfo(id, param, psize, buf, 0);
        if (CL_SUCCESS != res)
            throw std::runtime_error(std::string("clGetPlatformInfo failed"));

        // just in case, ensure trailing zero for ASCIIZ string
        buf[psize] = 0;

        paramStr = buf;

        return CL_SUCCESS;
    }

private:
    std::string m_profile;
    std::string m_version;
    std::string m_name;
    std::string m_vendor;
    std::string m_extensions;
};


class DeviceInfo
{
public:
    DeviceInfo()
    {}

    ~DeviceInfo()
    {}

    cl_int QueryInfo(cl_device_id id)
    {
        query_param(id, CL_DEVICE_TYPE, m_type);
        query_param(id, CL_DEVICE_VENDOR_ID, m_vendor_id);
        query_param(id, CL_DEVICE_MAX_COMPUTE_UNITS, m_max_compute_units);
        query_param(id, CL_DEVICE_MAX_WORK_ITEM_DIMENSIONS, m_max_work_item_dimensions);
        query_param(id, CL_DEVICE_MAX_WORK_ITEM_SIZES, m_max_work_item_sizes);
        query_param(id, CL_DEVICE_MAX_WORK_GROUP_SIZE, m_max_work_group_size);
        query_param(id, CL_DEVICE_PREFERRED_VECTOR_WIDTH_CHAR, m_preferred_vector_width_char);
        query_param(id, CL_DEVICE_PREFERRED_VECTOR_WIDTH_SHORT, m_preferred_vector_width_short);
        query_param(id, CL_DEVICE_PREFERRED_VECTOR_WIDTH_INT, m_preferred_vector_width_int);
        query_param(id, CL_DEVICE_PREFERRED_VECTOR_WIDTH_LONG, m_preferred_vector_width_long);
        query_param(id, CL_DEVICE_PREFERRED_VECTOR_WIDTH_FLOAT, m_preferred_vector_width_float);
        query_param(id, CL_DEVICE_PREFERRED_VECTOR_WIDTH_DOUBLE, m_preferred_vector_width_double);
#if defined(CL_VERSION_1_1)
        query_param(id, CL_DEVICE_PREFERRED_VECTOR_WIDTH_HALF, m_preferred_vector_width_half);
        query_param(id, CL_DEVICE_NATIVE_VECTOR_WIDTH_CHAR, m_native_vector_width_char);
        query_param(id, CL_DEVICE_NATIVE_VECTOR_WIDTH_SHORT, m_native_vector_width_short);
        query_param(id, CL_DEVICE_NATIVE_VECTOR_WIDTH_INT, m_native_vector_width_int);
        query_param(id, CL_DEVICE_NATIVE_VECTOR_WIDTH_LONG, m_native_vector_width_long);
        query_param(id, CL_DEVICE_NATIVE_VECTOR_WIDTH_FLOAT, m_native_vector_width_float);
        query_param(id, CL_DEVICE_NATIVE_VECTOR_WIDTH_DOUBLE, m_native_vector_width_double);
        query_param(id, CL_DEVICE_NATIVE_VECTOR_WIDTH_HALF, m_native_vector_width_half);
#endif
        query_param(id, CL_DEVICE_MAX_CLOCK_FREQUENCY, m_max_clock_frequency);
        query_param(id, CL_DEVICE_ADDRESS_BITS, m_address_bits);
        query_param(id, CL_DEVICE_MAX_MEM_ALLOC_SIZE, m_max_mem_alloc_size);
        query_param(id, CL_DEVICE_IMAGE_SUPPORT, m_image_support);
        query_param(id, CL_DEVICE_MAX_READ_IMAGE_ARGS, m_max_read_image_args);
        query_param(id, CL_DEVICE_MAX_WRITE_IMAGE_ARGS, m_max_write_image_args);
#if defined(CL_VERSION_2_0)
        query_param(id, CL_DEVICE_MAX_READ_WRITE_IMAGE_ARGS, m_max_read_write_image_args);
#endif
        query_param(id, CL_DEVICE_IMAGE2D_MAX_WIDTH, m_image2d_max_width);
        query_param(id, CL_DEVICE_IMAGE2D_MAX_HEIGHT, m_image2d_max_height);
        query_param(id, CL_DEVICE_IMAGE3D_MAX_WIDTH, m_image3d_max_width);
        query_param(id, CL_DEVICE_IMAGE3D_MAX_HEIGHT, m_image3d_max_height);
        query_param(id, CL_DEVICE_IMAGE3D_MAX_DEPTH, m_image3d_max_depth);
#if defined(CL_VERSION_1_2)
        query_param(id, CL_DEVICE_IMAGE_MAX_BUFFER_SIZE, m_image_max_buffer_size);
        query_param(id, CL_DEVICE_IMAGE_MAX_ARRAY_SIZE, m_image_max_array_size);
#endif
        query_param(id, CL_DEVICE_MAX_SAMPLERS, m_max_samplers);
#if defined(CL_VERSION_1_2)
        query_param(id, CL_DEVICE_IMAGE_PITCH_ALIGNMENT, m_image_pitch_alignment);
        query_param(id, CL_DEVICE_IMAGE_BASE_ADDRESS_ALIGNMENT, m_image_base_address_alignment);
#endif
#if defined(CL_VERSION_2_0)
        query_param(id, CL_DEVICE_MAX_PIPE_ARGS, m_max_pipe_args);
        query_param(id, CL_DEVICE_PIPE_MAX_ACTIVE_RESERVATIONS, m_pipe_max_active_reservations);
        query_param(id, CL_DEVICE_PIPE_MAX_PACKET_SIZE, m_pipe_max_packet_size);
#endif
        query_param(id, CL_DEVICE_MAX_PARAMETER_SIZE, m_max_parameter_size);
        query_param(id, CL_DEVICE_MEM_BASE_ADDR_ALIGN, m_mem_base_addr_align);
        query_param(id, CL_DEVICE_SINGLE_FP_CONFIG, m_single_fp_config);
#if defined(CL_VERSION_1_2)
        query_param(id, CL_DEVICE_DOUBLE_FP_CONFIG, m_double_fp_config);
#endif
        query_param(id, CL_DEVICE_GLOBAL_MEM_CACHE_TYPE, m_global_mem_cache_type);
        query_param(id, CL_DEVICE_GLOBAL_MEM_CACHELINE_SIZE, m_global_mem_cacheline_size);
        query_param(id, CL_DEVICE_GLOBAL_MEM_CACHE_SIZE, m_global_mem_cache_size);
        query_param(id, CL_DEVICE_GLOBAL_MEM_SIZE, m_global_mem_size);
        query_param(id, CL_DEVICE_MAX_CONSTANT_BUFFER_SIZE, m_max_constant_buffer_size);
        query_param(id, CL_DEVICE_MAX_CONSTANT_ARGS, m_max_constant_args);
#if defined(CL_VERSION_2_0)
        query_param(id, CL_DEVICE_MAX_GLOBAL_VARIABLE_SIZE, m_max_global_variable_size);
        query_param(id, CL_DEVICE_GLOBAL_VARIABLE_PREFERRED_TOTAL_SIZE, m_global_variable_preferred_total_size);
#endif
        query_param(id, CL_DEVICE_LOCAL_MEM_TYPE, m_local_mem_type);
        query_param(id, CL_DEVICE_LOCAL_MEM_SIZE, m_local_mem_size);
        query_param(id, CL_DEVICE_ERROR_CORRECTION_SUPPORT, m_error_correction_support);
#if defined(CL_VERSION_1_1)
        query_param(id, CL_DEVICE_HOST_UNIFIED_MEMORY, m_host_unified_memory);
#endif
        query_param(id, CL_DEVICE_PROFILING_TIMER_RESOLUTION, m_profiling_timer_resolution);
        query_param(id, CL_DEVICE_ENDIAN_LITTLE, m_endian_little);
        query_param(id, CL_DEVICE_AVAILABLE, m_available);
        query_param(id, CL_DEVICE_COMPILER_AVAILABLE, m_compiler_available);
#if defined(CL_VERSION_1_2)
        query_param(id, CL_DEVICE_LINKER_AVAILABLE, m_linker_available);
#endif
        query_param(id, CL_DEVICE_EXECUTION_CAPABILITIES, m_execution_capabilities);
        query_param(id, CL_DEVICE_QUEUE_PROPERTIES, m_queue_properties);
#if defined(CL_VERSION_2_0)
        query_param(id, CL_DEVICE_QUEUE_ON_HOST_PROPERTIES, m_queue_on_host_properties);
        query_param(id, CL_DEVICE_QUEUE_ON_DEVICE_PROPERTIES, m_queue_on_device_properties);
        query_param(id, CL_DEVICE_QUEUE_ON_DEVICE_PREFERRED_SIZE, m_queue_on_device_preferred_size);
        query_param(id, CL_DEVICE_QUEUE_ON_DEVICE_MAX_SIZE, m_queue_on_device_max_size);
        query_param(id, CL_DEVICE_MAX_ON_DEVICE_QUEUES, m_max_on_device_queues);
        query_param(id, CL_DEVICE_MAX_ON_DEVICE_EVENTS, m_max_on_device_events);
#endif
#if defined(CL_VERSION_1_2)
        query_param(id, CL_DEVICE_BUILT_IN_KERNELS, m_built_in_kernels);
#endif
        query_param(id, CL_DEVICE_PLATFORM, m_platform);
        query_param(id, CL_DEVICE_NAME, m_name);
        query_param(id, CL_DEVICE_VENDOR, m_vendor);
        query_param(id, CL_DRIVER_VERSION, m_driver_version);
        query_param(id, CL_DEVICE_PROFILE, m_profile);
        query_param(id, CL_DEVICE_VERSION, m_version);
#if defined(CL_VERSION_1_1)
        query_param(id, CL_DEVICE_OPENCL_C_VERSION, m_opencl_c_version);
#endif
        query_param(id, CL_DEVICE_EXTENSIONS, m_extensions);
#if defined(CL_VERSION_1_2)
        query_param(id, CL_DEVICE_PRINTF_BUFFER_SIZE, m_printf_buffer_size);
        query_param(id, CL_DEVICE_PREFERRED_INTEROP_USER_SYNC, m_preferred_interop_user_sync);
        query_param(id, CL_DEVICE_PARENT_DEVICE, m_parent_device);
        query_param(id, CL_DEVICE_PARTITION_MAX_SUB_DEVICES, m_partition_max_sub_devices);
        query_param(id, CL_DEVICE_PARTITION_PROPERTIES, m_partition_properties);
        query_param(id, CL_DEVICE_PARTITION_AFFINITY_DOMAIN, m_partition_affinity_domain);
        query_param(id, CL_DEVICE_PARTITION_TYPE, m_partition_type);
        query_param(id, CL_DEVICE_REFERENCE_COUNT, m_reference_count);
#endif
        return CL_SUCCESS;
    }

    std::string Name() { return m_name; }

private:
    template<typename T>
    cl_int query_param(cl_device_id id, cl_device_info param, T& value)
    {
        cl_int res;
        size_t size = 0;

        res = clGetDeviceInfo(id, param, 0, 0, &size);
        if (CL_SUCCESS != res && size != 0)
            throw std::runtime_error(std::string("clGetDeviceInfo failed"));

        if (0 == size)
            return CL_SUCCESS;

        if (sizeof(T) != size)
            throw std::runtime_error(std::string("clGetDeviceInfo: param size mismatch"));

        res = clGetDeviceInfo(id, param, size, &value, 0);
        if (CL_SUCCESS != res)
            throw std::runtime_error(std::string("clGetDeviceInfo failed"));

        return CL_SUCCESS;
    }

    template<typename T>
    cl_int query_param(cl_device_id id, cl_device_info param, std::vector<T>& value)
    {
        cl_int res;
        size_t size;

        res = clGetDeviceInfo(id, param, 0, 0, &size);
        if (CL_SUCCESS != res)
            throw std::runtime_error(std::string("clGetDeviceInfo failed"));

        if (0 == size)
            return CL_SUCCESS;

        value.resize(size / sizeof(T));

        res = clGetDeviceInfo(id, param, size, &value[0], 0);
        if (CL_SUCCESS != res)
            throw std::runtime_error(std::string("clGetDeviceInfo failed"));

        return CL_SUCCESS;
    }

    cl_int query_param(cl_device_id id, cl_device_info param, std::string& value)
    {
        cl_int res;
        size_t size;

        res = clGetDeviceInfo(id, param, 0, 0, &size);
        if (CL_SUCCESS != res)
            throw std::runtime_error(std::string("clGetDeviceInfo failed"));

        value.resize(size + 1);

        res = clGetDeviceInfo(id, param, size, &value[0], 0);
        if (CL_SUCCESS != res)
            throw std::runtime_error(std::string("clGetDeviceInfo failed"));

        // just in case, ensure trailing zero for ASCIIZ string
        value[size] = 0;

        return CL_SUCCESS;
    }

private:
    cl_device_type                            m_type;
    cl_uint                                   m_vendor_id;
    cl_uint                                   m_max_compute_units;
    cl_uint                                   m_max_work_item_dimensions;
    std::vector<size_t>                       m_max_work_item_sizes;
    size_t                                    m_max_work_group_size;
    cl_uint                                   m_preferred_vector_width_char;
    cl_uint                                   m_preferred_vector_width_short;
    cl_uint                                   m_preferred_vector_width_int;
    cl_uint                                   m_preferred_vector_width_long;
    cl_uint                                   m_preferred_vector_width_float;
    cl_uint                                   m_preferred_vector_width_double;
#if defined(CL_VERSION_1_1)
    cl_uint                                   m_preferred_vector_width_half;
    cl_uint                                   m_native_vector_width_char;
    cl_uint                                   m_native_vector_width_short;
    cl_uint                                   m_native_vector_width_int;
    cl_uint                                   m_native_vector_width_long;
    cl_uint                                   m_native_vector_width_float;
    cl_uint                                   m_native_vector_width_double;
    cl_uint                                   m_native_vector_width_half;
#endif
    cl_uint                                   m_max_clock_frequency;
    cl_uint                                   m_address_bits;
    cl_ulong                                  m_max_mem_alloc_size;
    cl_bool                                   m_image_support;
    cl_uint                                   m_max_read_image_args;
    cl_uint                                   m_max_write_image_args;
#if defined(CL_VERSION_2_0)
    cl_uint                                   m_max_read_write_image_args;
#endif
    size_t                                    m_image2d_max_width;
    size_t                                    m_image2d_max_height;
    size_t                                    m_image3d_max_width;
    size_t                                    m_image3d_max_height;
    size_t                                    m_image3d_max_depth;
#if defined(CL_VERSION_1_2)
    size_t                                    m_image_max_buffer_size;
    size_t                                    m_image_max_array_size;
#endif
    cl_uint                                   m_max_samplers;
#if defined(CL_VERSION_1_2)
    cl_uint                                   m_image_pitch_alignment;
    cl_uint                                   m_image_base_address_alignment;
#endif
#if defined(CL_VERSION_2_0)
    cl_uint                                   m_max_pipe_args;
    cl_uint                                   m_pipe_max_active_reservations;
    cl_uint                                   m_pipe_max_packet_size;
#endif
    size_t                                    m_max_parameter_size;
    cl_uint                                   m_mem_base_addr_align;
    cl_device_fp_config                       m_single_fp_config;
#if defined(CL_VERSION_1_2)
    cl_device_fp_config                       m_double_fp_config;
#endif
    cl_device_mem_cache_type                  m_global_mem_cache_type;
    cl_uint                                   m_global_mem_cacheline_size;
    cl_ulong                                  m_global_mem_cache_size;
    cl_ulong                                  m_global_mem_size;
    cl_ulong                                  m_max_constant_buffer_size;
    cl_uint                                   m_max_constant_args;
#if defined(CL_VERSION_2_0)
    size_t                                    m_max_global_variable_size;
    size_t                                    m_global_variable_preferred_total_size;
#endif
    cl_device_local_mem_type                  m_local_mem_type;
    cl_ulong                                  m_local_mem_size;
    cl_bool                                   m_error_correction_support;
#if defined(CL_VERSION_1_1)
    cl_bool                                   m_host_unified_memory;
#endif
    size_t                                    m_profiling_timer_resolution;
    cl_bool                                   m_endian_little;
    cl_bool                                   m_available;
    cl_bool                                   m_compiler_available;
#if defined(CL_VERSION_1_2)
    cl_bool                                   m_linker_available;
#endif
    cl_device_exec_capabilities               m_execution_capabilities;
    cl_command_queue_properties               m_queue_properties;
#if defined(CL_VERSION_2_0)
    cl_command_queue_properties               m_queue_on_host_properties;
    cl_command_queue_properties               m_queue_on_device_properties;
    cl_uint                                   m_queue_on_device_preferred_size;
    cl_uint                                   m_queue_on_device_max_size;
    cl_uint                                   m_max_on_device_queues;
    cl_uint                                   m_max_on_device_events;
#endif
#if defined(CL_VERSION_1_2)
    std::string                               m_built_in_kernels;
#endif
    cl_platform_id                            m_platform;
    std::string                               m_name;
    std::string                               m_vendor;
    std::string                               m_driver_version;
    std::string                               m_profile;
    std::string                               m_version;
#if defined(CL_VERSION_1_1)
    std::string                               m_opencl_c_version;
#endif
    std::string                               m_extensions;
#if defined(CL_VERSION_1_2)
    size_t                                    m_printf_buffer_size;
    cl_bool                                   m_preferred_interop_user_sync;
    cl_device_id                              m_parent_device;
    cl_uint                                   m_partition_max_sub_devices;
    std::vector<cl_device_partition_property> m_partition_properties;
    cl_device_affinity_domain                 m_partition_affinity_domain;
    std::vector<cl_device_partition_property> m_partition_type;
    cl_uint                                   m_reference_count;
#endif
};

} // namespace opencl


class App
{
public:
    App(CommandLineParser& cmd);
    ~App();

    int initOpenCL();
    int initVideoSource();

    int process_frame_with_open_cl(cv::Mat& frame, bool use_buffer, cl_mem* cl_buffer);
    int process_cl_buffer_with_opencv(cl_mem buffer, size_t step, int rows, int cols, int type, cv::UMat& u);
    int process_cl_image_with_opencv(cl_mem image, cv::UMat& u);

    int run();

    bool isRunning() { return m_running; }
    bool doProcess() { return m_process; }
    bool useBuffer() { return m_use_buffer; }

    void setRunning(bool running)      { m_running = running; }
    void setDoProcess(bool process)    { m_process = process; }
    void setUseBuffer(bool use_buffer) { m_use_buffer = use_buffer; }

protected:
    bool nextFrame(cv::Mat& frame) { return m_cap.read(frame); }
    void handleKey(char key);
    void timerStart();
    void timerEnd();
    std::string timeStr() const;
    std::string message() const;

private:
    bool                        m_running;
    bool                        m_process;
    bool                        m_use_buffer;

    int64                       m_t0;
    int64                       m_t1;
    float                       m_time;
    float                       m_frequency;

    string                      m_file_name;
    int                         m_camera_id;
    cv::VideoCapture            m_cap;
    cv::Mat                     m_frame;
    cv::Mat                     m_frameGray;

    opencl::PlatformInfo        m_platformInfo;
    opencl::DeviceInfo          m_deviceInfo;
    std::vector<cl_platform_id> m_platform_ids;
    cl_context                  m_context;
    cl_device_id                m_device_id;
    cl_command_queue            m_queue;
    cl_program                  m_program;
    cl_kernel                   m_kernelBuf;
    cl_kernel                   m_kernelImg;
    cl_mem                      m_img_src; // used as src in case processing of cl image
    cl_mem                      m_mem_obj;
    cl_event                    m_event;
};


App::App(CommandLineParser& cmd)
{
    cout << "
Press ESC to exit
" << endl;
    cout << "
      'p' to toggle ON/OFF processing
" << endl;
    cout << "
       SPACE to switch between OpenCL buffer/image
" << endl;

    m_camera_id  = cmd.get<int>("camera");
    m_file_name  = cmd.get<string>("video");

    m_running    = false;
    m_process    = false;
    m_use_buffer = false;

    m_t0         = 0;
    m_t1         = 0;
    m_time       = 0.0;
    m_frequency  = (float)cv::getTickFrequency();

    m_context    = 0;
    m_device_id  = 0;
    m_queue      = 0;
    m_program    = 0;
    m_kernelBuf  = 0;
    m_kernelImg  = 0;
    m_img_src    = 0;
    m_mem_obj    = 0;
    m_event      = 0;
} // ctor


App::~App()
{
    if (m_queue)
    {
        clFinish(m_queue);
        clReleaseCommandQueue(m_queue);
        m_queue = 0;
    }

    if (m_program)
    {
        clReleaseProgram(m_program);
        m_program = 0;
    }

    if (m_img_src)
    {
        clReleaseMemObject(m_img_src);
        m_img_src = 0;
    }

    if (m_mem_obj)
    {
        clReleaseMemObject(m_mem_obj);
        m_mem_obj = 0;
    }

    if (m_event)
    {
        clReleaseEvent(m_event);
    }

    if (m_kernelBuf)
    {
        clReleaseKernel(m_kernelBuf);
        m_kernelBuf = 0;
    }

    if (m_kernelImg)
    {
        clReleaseKernel(m_kernelImg);
        m_kernelImg = 0;
    }

    if (m_device_id)
    {
        clReleaseDevice(m_device_id);
        m_device_id = 0;
    }

    if (m_context)
    {
        clReleaseContext(m_context);
        m_context = 0;
    }
} // dtor


int App::initOpenCL()
{
    cl_int res = CL_SUCCESS;
    cl_uint num_entries = 0;

    res = clGetPlatformIDs(0, 0, &num_entries);
    if (CL_SUCCESS != res)
        return -1;

    m_platform_ids.resize(num_entries);

    res = clGetPlatformIDs(num_entries, &m_platform_ids[0], 0);
    if (CL_SUCCESS != res)
        return -1;

    unsigned int i;

    // create context from first platform with GPU device
    for (i = 0; i < m_platform_ids.size(); i++)
    {
        cl_context_properties props[] =
        {
            CL_CONTEXT_PLATFORM,
            (cl_context_properties)(m_platform_ids[i]),
            0
        };

        m_context = clCreateContextFromType(props, CL_DEVICE_TYPE_GPU, 0, 0, &res);
        if (0 == m_context || CL_SUCCESS != res)
            continue;

        res = clGetContextInfo(m_context, CL_CONTEXT_DEVICES, sizeof(cl_device_id), &m_device_id, 0);
        if (CL_SUCCESS != res)
            return -1;

        m_queue = clCreateCommandQueue(m_context, m_device_id, 0, &res);
        if (0 == m_queue || CL_SUCCESS != res)
            return -1;

        const char* kernelSrc =
            "__kernel "
            "void bitwise_inv_buf_8uC1("
            "    __global unsigned char* pSrcDst,"
            "             int            srcDstStep,"
            "             int            rows,"
            "             int            cols)"
            "{"
            "    int x = get_global_id(0);"
            "    int y = get_global_id(1);"
            "    int idx = mad24(y, srcDstStep, x);"
            "    pSrcDst[idx] = ~pSrcDst[idx];"
            "}"
            "__kernel "
            "void bitwise_inv_img_8uC1("
            "    read_only  image2d_t srcImg,"
            "    write_only image2d_t dstImg)"
            "{"
            "    int x = get_global_id(0);"
            "    int y = get_global_id(1);"
            "    int2 coord = (int2)(x, y);"
            "    uint4 val = read_imageui(srcImg, coord);"
            "    val.x = (~val.x) & 0x000000FF;"
            "    write_imageui(dstImg, coord, val);"
            "}";
        size_t len = strlen(kernelSrc);
        m_program = clCreateProgramWithSource(m_context, 1, &kernelSrc, &len, &res);
        if (0 == m_program || CL_SUCCESS != res)
            return -1;

        res = clBuildProgram(m_program, 1, &m_device_id, 0, 0, 0);
        if (CL_SUCCESS != res)
            return -1;

        m_kernelBuf = clCreateKernel(m_program, "bitwise_inv_buf_8uC1", &res);
        if (0 == m_kernelBuf || CL_SUCCESS != res)
            return -1;

        m_kernelImg = clCreateKernel(m_program, "bitwise_inv_img_8uC1", &res);
        if (0 == m_kernelImg || CL_SUCCESS != res)
            return -1;

        m_platformInfo.QueryInfo(m_platform_ids[i]);
        m_deviceInfo.QueryInfo(m_device_id);

        // attach OpenCL context to OpenCV
        cv::ocl::attachContext(m_platformInfo.Name(), m_platform_ids[i], m_context, m_device_id);

        break;
    }

    return m_context != 0 ? CL_SUCCESS : -1;
} // initOpenCL()


int App::initVideoSource()
{
    try
    {
        if (!m_file_name.empty() && m_camera_id == -1)
        {
            m_cap.open(m_file_name.c_str());
            if (!m_cap.isOpened())
                throw std::runtime_error(std::string("can't open video file: " + m_file_name));
        }
        else if (m_camera_id != -1)
        {
            m_cap.open(m_camera_id);
            if (!m_cap.isOpened())
            {
                std::stringstream msg;
                msg << "can't open camera: " << m_camera_id;
                throw std::runtime_error(msg.str());
            }
        }
        else
            throw std::runtime_error(std::string("specify video source"));
    }

    catch (std::exception e)
    {
        cerr << "ERROR: " << e.what() << std::endl;
        return -1;
    }

    return 0;
} // initVideoSource()


// this function is an example of "typical" OpenCL processing pipeline
// It creates OpenCL buffer or image, depending on use_buffer flag,
// from input media frame and process these data
// (inverts each pixel value in half of frame) with OpenCL kernel
int App::process_frame_with_open_cl(cv::Mat& frame, bool use_buffer, cl_mem* mem_obj)
{
    cl_int res = CL_SUCCESS;

    CV_Assert(mem_obj);

    cl_kernel kernel = 0;
    cl_mem mem = mem_obj[0];

    if (0 == mem || 0 == m_img_src)
    {
        // allocate/delete cl memory objects every frame for the simplicity.
        // in real applicaton more efficient pipeline can be built.

        if (use_buffer)
        {
            cl_mem_flags flags = CL_MEM_READ_WRITE | CL_MEM_USE_HOST_PTR;

            mem = clCreateBuffer(m_context, flags, frame.total(), frame.ptr(), &res);
            if (0 == mem || CL_SUCCESS != res)
                return -1;

            res = clSetKernelArg(m_kernelBuf, 0, sizeof(cl_mem), &mem);
            if (CL_SUCCESS != res)
                return -1;

            res = clSetKernelArg(m_kernelBuf, 1, sizeof(int), &frame.step[0]);
            if (CL_SUCCESS != res)
                return -1;

            res = clSetKernelArg(m_kernelBuf, 2, sizeof(int), &frame.rows);
            if (CL_SUCCESS != res)
                return -1;

            int cols2 = frame.cols / 2;
            res = clSetKernelArg(m_kernelBuf, 3, sizeof(int), &cols2);
            if (CL_SUCCESS != res)
                return -1;

            kernel = m_kernelBuf;
        }
        else
        {
            cl_mem_flags flags_src = CL_MEM_READ_ONLY | CL_MEM_USE_HOST_PTR;

            cl_image_format fmt;
            fmt.image_channel_order     = CL_R;
            fmt.image_channel_data_type = CL_UNSIGNED_INT8;

            cl_image_desc desc_src;
            desc_src.image_type        = CL_MEM_OBJECT_IMAGE2D;
            desc_src.image_width       = frame.cols;
            desc_src.image_height      = frame.rows;
            desc_src.image_depth       = 0;
            desc_src.image_array_size  = 0;
            desc_src.image_row_pitch   = frame.step[0];
            desc_src.image_slice_pitch = 0;
            desc_src.num_mip_levels    = 0;
            desc_src.num_samples       = 0;
            desc_src.buffer            = 0;
            m_img_src = clCreateImage(m_context, flags_src, &fmt, &desc_src, frame.ptr(), &res);
            if (0 == m_img_src || CL_SUCCESS != res)
                return -1;

            cl_mem_flags flags_dst = CL_MEM_READ_WRITE | CL_MEM_ALLOC_HOST_PTR;

            cl_image_desc desc_dst;
            desc_dst.image_type        = CL_MEM_OBJECT_IMAGE2D;
            desc_dst.image_width       = frame.cols;
            desc_dst.image_height      = frame.rows;
            desc_dst.image_depth       = 0;
            desc_dst.image_array_size  = 0;
            desc_dst.image_row_pitch   = 0;
            desc_dst.image_slice_pitch = 0;
            desc_dst.num_mip_levels    = 0;
            desc_dst.num_samples       = 0;
            desc_dst.buffer            = 0;
            mem = clCreateImage(m_context, flags_dst, &fmt, &desc_dst, 0, &res);
            if (0 == mem || CL_SUCCESS != res)
                return -1;

            size_t origin[] = { 0, 0, 0 };
            size_t region[] = { (size_t)frame.cols, (size_t)frame.rows, 1 };
            res = clEnqueueCopyImage(m_queue, m_img_src, mem, origin, origin, region, 0, 0, &m_event);
            if (CL_SUCCESS != res)
                return -1;

            res = clWaitForEvents(1, &m_event);
            if (CL_SUCCESS != res)
                return -1;

            res = clSetKernelArg(m_kernelImg, 0, sizeof(cl_mem), &m_img_src);
            if (CL_SUCCESS != res)
                return -1;

            res = clSetKernelArg(m_kernelImg, 1, sizeof(cl_mem), &mem);
            if (CL_SUCCESS != res)
                return -1;

            kernel = m_kernelImg;
        }
    }

    m_event = clCreateUserEvent(m_context, &res);
    if (0 == m_event || CL_SUCCESS != res)
        return -1;

    // process left half of frame in OpenCL
    size_t size[] = { (size_t)frame.cols / 2, (size_t)frame.rows };
    res = clEnqueueNDRangeKernel(m_queue, kernel, 2, 0, size, 0, 0, 0, &m_event);
    if (CL_SUCCESS != res)
        return -1;

    res = clWaitForEvents(1, &m_event);
    if (CL_SUCCESS != res)
        return - 1;

    mem_obj[0] = mem;

    return  0;
}


// this function is an example of interoperability between OpenCL buffer
// and OpenCV UMat objects. It converts (without copying data) OpenCL buffer
// to OpenCV UMat and then do blur on these data
int App::process_cl_buffer_with_opencv(cl_mem buffer, size_t step, int rows, int cols, int type, cv::UMat& u)
{
    cv::ocl::convertFromBuffer(buffer, step, rows, cols, type, u);

    // process right half of frame in OpenCV
    cv::Point pt(u.cols / 2, 0);
    cv::Size  sz(u.cols / 2, u.rows);
    cv::Rect roi(pt, sz);
    cv::UMat uroi(u, roi);
    cv::blur(uroi, uroi, cv::Size(7, 7), cv::Point(-3, -3));

    if (buffer)
        clReleaseMemObject(buffer);
    m_mem_obj = 0;

    return 0;
}


// this function is an example of interoperability between OpenCL image
// and OpenCV UMat objects. It converts OpenCL image
// to OpenCV UMat and then do blur on these data
int App::process_cl_image_with_opencv(cl_mem image, cv::UMat& u)
{
    cv::ocl::convertFromImage(image, u);

    // process right half of frame in OpenCV
    cv::Point pt(u.cols / 2, 0);
    cv::Size  sz(u.cols / 2, u.rows);
    cv::Rect roi(pt, sz);
    cv::UMat uroi(u, roi);
    cv::blur(uroi, uroi, cv::Size(7, 7), cv::Point(-3, -3));

    if (image)
        clReleaseMemObject(image);
    m_mem_obj = 0;

    if (m_img_src)
        clReleaseMemObject(m_img_src);
    m_img_src = 0;

    return 0;
}


int App::run()
{
    if (0 != initOpenCL())
        return -1;

    if (0 != initVideoSource())
        return -1;

    Mat img_to_show;

    // set running state until ESC pressed
    setRunning(true);
    // set process flag to show some data processing
    // can be toggled on/off by 'p' button
    setDoProcess(true);
    // set use buffer flag,
    // when it is set to true, will demo interop opencl buffer and cv::Umat,
    // otherwise demo interop opencl image and cv::UMat
    // can be switched on/of by SPACE button
    setUseBuffer(true);

    // Iterate over all frames
    while (isRunning() && nextFrame(m_frame))
    {
        cv::cvtColor(m_frame, m_frameGray, COLOR_BGR2GRAY);

        UMat uframe;

        // work
        timerStart();

        if (doProcess())
        {
            process_frame_with_open_cl(m_frameGray, useBuffer(), &m_mem_obj);

            if (useBuffer())
                process_cl_buffer_with_opencv(
                    m_mem_obj, m_frameGray.step[0], m_frameGray.rows, m_frameGray.cols, m_frameGray.type(), uframe);
            else
                process_cl_image_with_opencv(m_mem_obj, uframe);
        }
        else
        {
            m_frameGray.copyTo(uframe);
        }

        timerEnd();

        uframe.copyTo(img_to_show);

        putText(img_to_show, "Version : " + m_platformInfo.Version(), Point(5, 30), FONT_HERSHEY_SIMPLEX, 1., Scalar(255, 100, 0), 2);
        putText(img_to_show, "Name : " + m_platformInfo.Name(), Point(5, 60), FONT_HERSHEY_SIMPLEX, 1., Scalar(255, 100, 0), 2);
        putText(img_to_show, "Device : " + m_deviceInfo.Name(), Point(5, 90), FONT_HERSHEY_SIMPLEX, 1., Scalar(255, 100, 0), 2);
        cv::String memtype = useBuffer() ? "buffer" : "image";
        putText(img_to_show, "interop with OpenCL " + memtype, Point(5, 120), FONT_HERSHEY_SIMPLEX, 1., Scalar(255, 100, 0), 2);
        putText(img_to_show, "Time : " + timeStr() + " msec", Point(5, 150), FONT_HERSHEY_SIMPLEX, 1., Scalar(255, 100, 0), 2);

        imshow("opencl_interop", img_to_show);

        handleKey((char)waitKey(3));
    }

    return 0;
}


void App::handleKey(char key)
{
    switch (key)
    {
    case 27:
        setRunning(false);
        break;

    case ' ':
        setUseBuffer(!useBuffer());
        break;

    case 'p':
    case 'P':
        setDoProcess( !doProcess() );
        break;

    default:
        break;
    }
}


inline void App::timerStart()
{
    m_t0 = getTickCount();
}


inline void App::timerEnd()
{
    m_t1 = getTickCount();
    int64 delta = m_t1 - m_t0;
    m_time = (delta / m_frequency) * 1000; // units msec
}


inline string App::timeStr() const
{
    stringstream ss;
    ss << std::fixed << std::setprecision(1) << m_time;
    return ss.str();
}


int main(int argc, char** argv)
{
    const char* keys =
        "{ help h ?    |          | print help message }"
        "{ camera c    | -1       | use camera as input }"
        "{ video  v    |          | use video as input }";

    CommandLineParser cmd(argc, argv, keys);
    if (cmd.has("help"))
    {
        cmd.printMessage();
        return EXIT_SUCCESS;
    }

    App app(cmd);

    try
    {
        app.run();
    }

    catch (const cv::Exception& e)
    {
        cout << "error: " << e.what() << endl;
        return 1;
    }

    catch (const std::exception& e)
    {
        cout << "error: " << e.what() << endl;
        return 1;
    }

    catch (...)
    {
        cout << "unknown exception" << endl;
        return 1;
    }

    return EXIT_SUCCESS;
} // main()
4.2.2 Makefile
TARGET = main
CXX = g++
CFLAGS += -I/usr/include -I/usr/local/include/opencv -I/usr/local/include/opencv2 -L/usr/lib  -L/usr/local/lib -L/lib -std=c++98

CFLAGS +=  -lopencv_core -lopencv_objdetect -lopencv_highgui -lopencv_videoio -lopencv_imgcodecs -lopencv_imgproc -lOpenCL -lpthread -lrt

all:
    @$(CXX)  $(TARGET).cpp -o $(TARGET) $(CFLAGS)
clean:
    rm -rf  $(TARGET)
4.2.3 编译运行
root@NanoPC-T6:/opt/opencl-project/opencv-ocl# make 
root@NanoPC-T6:/opt/opencl-project/opencv-ocl# ./main -c

如下图所示:

参考文章

[1] RK3588实战:调用npu加速,yolov5识别图像、ffmpeg发送到rtmp服务器

[2] 嵌入式AI应用开发实战指南—基于LubanCat-RK系列板卡

[3] RK3588边缘计算

[4] OpenCL学习笔记(四)手动编译开发库(ubuntu+gcc+rk3588)

[5] 如何在RK3399中使用opencl并安装QT开发

[6] Arm Mali GPU OpenCL Developer Guide

[7] 什么是OpenCL

[8] 高性能计算

[9] OpenCL练习(一):使用OpenCL+OpenCV进行RGB转灰度图

[10] https://opencv.org/opencl

[11] https://github.com/opencv/opencv/wiki/OpenCL-optimizations

[12] OpenCV OpenCL support

[13] 一、Opencv-OCL编程基础