【Altera SoC体验之旅】+ 正式开启OpenCL模式

dongliudongliu

本文转自：http://bbs.eeworld.com.cn/thread-455862-1-1.html

最近可谓几经周折。先前的Lark板子虽然看上去很高端，但实在是资料太少，对于我的应用来说从头开始搭模块不太现实。
与EEWorld 影子 沟通后，在她帮助下，和网友 @chenzhufly 互换了板子，他用的是Arrow  SoC。这个板子资料丰富一些，至少在RocketBoard上有很多教程和资料。
一切看上去都很完美，但做完所有实验后发现，本来Altera承诺的“支持OpenCL开发”结果是一句口号，我找遍了官网也没有发现这块板子的BSP。问过了Arrow的员工 @Alex，得到回答也是暂时还没有BSP。
于是不得以，又换了一块支持OpenCL开发的板子——友晶的DE1-SoC，这块性价比最高的板子。与我交换板子的是 @coyoo 大神（《深入理解Altera FPGA应用设计》作者），不得不说，论坛果然卧虎藏龙啊。

有幸参加这次比赛，有幸体验了三块不同的板子（总共才4块，太值了），有幸认识了一群技术上的大牛，想想这次赚大发了。

一定有同学会问，你到底要做什么东东，非要用Open CL？

不止一个人问过这个问题了，其实我看到这个比赛时，想想自己都已经不是学生了，没有那么多课外时间搞比赛，所以没打算报名，但刚好看到在全球计算机大会上Altera与百度合作研发的深度神经网络加速器（DNN by FPGA），而自己恰好又有个想法在FPGA上完成卷积神经网络的搭建（工作相关），各种机缘巧合下，毅然报名了。

神经网络有什么用途？它是模拟人大脑的组织形式，用大量神经元之间相互传递消息实现认知功能的，最简单的例子就是物体识别，人看到一张桌子，就会知道这是个桌子，而不是凳子，因为符合“桌子”特征。在人脑中已经通过大量训练，将“桌子”特征记录在神经元之间的权值上了。而对于计算机，通过摄像头看到桌子时，只是一堆像素值（RGB），浅层次的处理如中值滤波，相关，Sobel滤波是无法认知“桌子”这个特征的，而只是将某一维度的信息呈现给用户，让用户自己判断。为了将信息有效组织，需要构建大量的相同功能的神经元，每个单元执行最基本的操作（将输入累加，满足条件时输出给下一个神经元），这样层层累积，最终实现深层次的认知功能，在最末端的神经元直接可以回答“这是个桌子”或者“这是个凳子”或者“这是个椅子”。
卷积神经网络是在上面神经网络基础上做了一些近似。将同一层的神经元权值共享，减少了连接数，有利于计算机实现。

好了，说了这么多，其实说白了一句话就是，我目前算法是用C/C++以及CUDA实现的，如果迁移到FPGA上运行，使用OpenCL是最快的方式，也是这次体验最重要的内容（以前在FPGA上开发都是VHDL/Verilog，设计+仿真验证+调试太花时间，短期内难以完成，而且我目前只关心算法，不关心底层实现，如果能实现最基本的功能，这一阶段就算完成了，后面再考虑资源、时序、性能上的优化。

拿到板子后，仔细阅读了官方文档，搭建OpenCL环境。

今天时间关系，不再详细展开OpenCL的语法、结构，直接上例子。

烧写TF卡，流程参考我之前的帖子。烧写完成，将SW10拨码开关设置为“01010”（这个很重要，如果没有配置FPGA，后面脚本会lock），上电启动。
上一张图：

PC上打开Putty，设置波特率115200，用户名root，没有密码，进入系统。

可以看得出系统是Poky 8.0 (Yocto Project 1.3 Reference Distro) 1.3 socfpga ttyS0，和之前Lark板子上默认的系统是一样的。
ls一下，当前目录下有很多例程。
先做个准备活动：运行初始化OpenCL环境的脚本：
source ./init_opencl.sh
很快就结束了。我们打开看下这个脚本内容都是什么东东？
root@socfpga:~/vector_Add# cat ~/init_opencl.sh
export ALTERAOCLSDKROOT=/home/root/opencl_arm32_rte
export AOCL_BOARD_PACKAGE_ROOT=$ALTERAOCLSDKROOT/board/c5soc
export PATH=$ALTERAOCLSDKROOT/bin:$PATH
export LD_LIBRARY_PATH=$ALTERAOCLSDKROOT/host/arm32/lib:$LD_LIBRARY_PATH
insmod $AOCL_BOARD_PACKAGE_ROOT/driver/aclsoc_drv.ko

首先设置了几个环境变量：
ALTERAOCLSDKROOT
AOCL_BOARD_PACKAGE_ROOT
PATH
LD_LIBRARY_PATH
之后执行了insmod操作，加载驱动。
我们可以知道OpenCL的服务是由驱动模块$AOCL_BOARD_PACKAGE_ROOT/driver/aclsoc_drv.ko 提供的。
OK，就绪，下面先进入helloworld目录。
root@socfpga:~# cd helloworld/
root@socfpga:~/helloworld# ls
hello_world.aocx  helloworld

这个目录有hello_world.aocx和 helloworld两个文件。前者运行在FPGA上（OpenCL中称为核函数, Kernel），后者运行在ARM上（OpenCL中称为主机程序，Host Program）。两者编译过程如图所示。

运行步骤如下：
root@socfpga:~/helloworld# aocl program /dev/acl0 hello_world.aocx
aocl program: Running reprogram from /home/root/opencl_arm32_rte/board/c5soc/arm32/bin
Reprogramming was successful!
root@socfpga:~/helloworld# ./helloworld
Querying platform for info:
==========================
CL_PLATFORM_NAME                         = Altera SDK for OpenCL
CL_PLATFORM_VENDOR                       = Altera Corporation
CL_PLATFORM_VERSION                      = OpenCL 1.0 Altera SDK for OpenCL, Version 14.0

Querying device for info:
========================
CL_DEVICE_NAME                           = de1soc_sharedonly : Cyclone V SoC Development Kit
CL_DEVICE_VENDOR                         = Altera Corporation
CL_DEVICE_VENDOR_ID                      = 4466
CL_DEVICE_VERSION                        = OpenCL 1.0 Altera SDK for OpenCL, Version 14.0
CL_DRIVER_VERSION                        = 14.0
CL_DEVICE_ADDRESS_BITS                   = 64
CL_DEVICE_AVAILABLE                      = true
CL_DEVICE_ENDIAN_LITTLE                  = true
CL_DEVICE_GLOBAL_MEM_CACHE_SIZE          = 32768
CL_DEVICE_GLOBAL_MEM_CACHELINE_SIZE      = 0
CL_DEVICE_GLOBAL_MEM_SIZE                = 536870912
CL_DEVICE_IMAGE_SUPPORT                  = false
CL_DEVICE_LOCAL_MEM_SIZE                 = 16384
CL_DEVICE_MAX_CLOCK_FREQUENCY            = 1000
CL_DEVICE_MAX_COMPUTE_UNITS              = 1
CL_DEVICE_MAX_CONSTANT_ARGS              = 8
CL_DEVICE_MAX_CONSTANT_BUFFER_SIZE       = 134217728
CL_DEVICE_MAX_WORK_ITEM_DIMENSIONS       = 3
CL_DEVICE_MAX_WORK_ITEM_DIMENSIONS       = 8192
CL_DEVICE_MIN_DATA_TYPE_ALIGN_SIZE       = 1024
CL_DEVICE_PREFERRED_VECTOR_WIDTH_CHAR    = 4
CL_DEVICE_PREFERRED_VECTOR_WIDTH_SHORT   = 2
CL_DEVICE_PREFERRED_VECTOR_WIDTH_INT     = 1
CL_DEVICE_PREFERRED_VECTOR_WIDTH_LONG    = 1
CL_DEVICE_PREFERRED_VECTOR_WIDTH_FLOAT   = 1
CL_DEVICE_PREFERRED_VECTOR_WIDTH_DOUBLE  = 0
Command queue out of order?              = false
Command queue profiling enabled?         = true
Using AOCX: hello_world.aocx

Kernel initialization is complete.
Launching the kernel...

Thread #2: Hello from Altera's OpenCL Compiler!

Kernel execution is complete.



可见，运行成功了。
想看源代码，可以在DE1-SoC_openCL_BSP.zip中找到，路径为examples/helloworld/。
后缀为.cl的文件为核函数。上面例子的核函数如下：
// AOC kernel demonstrating device-side printf call
__kernel void hello_world(int thread_id_from_which_to_print_message) {
  // Get index of the work item
  unsigned thread_id = get_global_id(0);

  if(thread_id == thread_id_from_which_to_print_message) {
    printf("Thread #%u: Hello from Altera's OpenCL Compiler!\n", thread_id);
  }
}

类似C函数，只不过前缀加上“__kernel”关键词，指定它运行在设备（FPGA）上。使用Altera的OpenCL工具就可以编译为FPGA比特流配置文件。
这里的函数功能很简单，只是判断自身线程号是否与主机指定的相同，如果相同则输出一句话，其他线程保持沉默。

（未完，跟帖中）

dongliudongliu

接着看下Host Program长什么样。
#include <assert.h>
#include <stdio.h>
#include <stdlib.h>
#include <math.h>
#include <cstring>
#include "CL/opencl.h"
#include "AOCL_Utils.h"

using namespace aocl_utils;

#define STRING_BUFFER_LEN 1024

// Runtime constants
// Used to define the work set over which this kernel will execute.
static const size_t work_group_size = 8;  // 8 threads in the demo workgroup
// Defines kernel argument value, which is the workitem ID that will
// execute a printf call
static const int thread_id_to_output = 2;

// OpenCL runtime configuration
static cl_platform_id platform = NULL;
static cl_device_id device = NULL;
static cl_context context = NULL;
static cl_command_queue queue = NULL;
static cl_kernel kernel = NULL;
static cl_program program = NULL;

// Function prototypes
bool init();
void cleanup();
static void device_info_ulong( cl_device_id device, cl_device_info param, const char* name);
static void device_info_uint( cl_device_id device, cl_device_info param, const char* name);
static void device_info_bool( cl_device_id device, cl_device_info param, const char* name);
static void device_info_string( cl_device_id device, cl_device_info param, const char* name);
static void display_device_info( cl_device_id device );

// Entry point.
int main() {
  cl_int status;

  if(!init()) {
    return -1;
  }

  // Set the kernel argument (argument 0)
  status = clSetKernelArg(kernel, 0, sizeof(cl_int), (void*)&thread_id_to_output);
  checkError(status, "Failed to set kernel arg 0");

  printf("\nKernel initialization is complete.\n");
  printf("Launching the kernel...\n\n");

  // Configure work set over which the kernel will execute
  size_t wgSize[3] = {work_group_size, 1, 1};
  size_t gSize[3] = {work_group_size, 1, 1};

  // Launch the kernel
  status = clEnqueueNDRangeKernel(queue, kernel, 1, NULL, gSize, wgSize, 0, NULL, NULL);
  checkError(status, "Failed to launch kernel");

  // Wait for command queue to complete pending events
  status = clFinish(queue);
  checkError(status, "Failed to finish");

  printf("\nKernel execution is complete.\n");

  // Free the resources allocated
  cleanup();

  return 0;
}

/////// HELPER FUNCTIONS ///////

bool init() {
  cl_int status;

  if(!setCwdToExeDir()) {
    return false;
  }

  // Get the OpenCL platform.
  platform = findPlatform("Altera");
  if(platform == NULL) {
    printf("ERROR: Unable to find Altera OpenCL platform.\n");
    return false;
  }

  // User-visible output - Platform information
  {
    char char_buffer[STRING_BUFFER_LEN];
    printf("Querying platform for info:\n");
    printf("==========================\n");
    clGetPlatformInfo(platform, CL_PLATFORM_NAME, STRING_BUFFER_LEN, char_buffer, NULL);
    printf("%-40s = %s\n", "CL_PLATFORM_NAME", char_buffer);
    clGetPlatformInfo(platform, CL_PLATFORM_VENDOR, STRING_BUFFER_LEN, char_buffer, NULL);
    printf("%-40s = %s\n", "CL_PLATFORM_VENDOR ", char_buffer);
    clGetPlatformInfo(platform, CL_PLATFORM_VERSION, STRING_BUFFER_LEN, char_buffer, NULL);
    printf("%-40s = %s\n\n", "CL_PLATFORM_VERSION ", char_buffer);
  }

  // Query the available OpenCL devices.
  scoped_array<cl_device_id> devices;
  cl_uint num_devices;

  devices.reset(getDevices(platform, CL_DEVICE_TYPE_ALL, &num_devices));

  // We'll just use the first device.
  device = devices[0];

  // Display some device information.
  display_device_info(device);

  // Create the context.
  context = clCreateContext(NULL, 1, &device, NULL, NULL, &status);
  checkError(status, "Failed to create context");

  // Create the command queue.
  queue = clCreateCommandQueue(context, device, CL_QUEUE_PROFILING_ENABLE, &status);
  checkError(status, "Failed to create command queue");

  // Create the program.
  std::string binary_file = getBoardBinaryFile("hello_world", device);
  printf("Using AOCX: %s\n", binary_file.c_str());
  program = createProgramFromBinary(context, binary_file.c_str(), &device, 1);

  // Build the program that was just created.
  status = clBuildProgram(program, 0, NULL, "", NULL, NULL);
  checkError(status, "Failed to build program");

  // Create the kernel - name passed in here must match kernel name in the
  // original CL file, that was compiled into an AOCX file using the AOC tool
  const char *kernel_name = "hello_world";  // Kernel name, as defined in the CL file
  kernel = clCreateKernel(program, kernel_name, &status);
  checkError(status, "Failed to create kernel");

  return true;
}

// Free the resources allocated during initialization
void cleanup() {
  if(kernel) {
    clReleaseKernel(kernel);  
  }
  if(program) {
    clReleaseProgram(program);
  }
  if(queue) {
    clReleaseCommandQueue(queue);
  }
  if(context) {
    clReleaseContext(context);
  }
}

// Helper functions to display parameters returned by OpenCL queries
static void device_info_ulong( cl_device_id device, cl_device_info param, const char* name) {
   cl_ulong a;
   clGetDeviceInfo(device, param, sizeof(cl_ulong), &a, NULL);
   printf("%-40s = %lu\n", name, a);
}
static void device_info_uint( cl_device_id device, cl_device_info param, const char* name) {
   cl_uint a;
   clGetDeviceInfo(device, param, sizeof(cl_uint), &a, NULL);
   printf("%-40s = %u\n", name, a);
}
static void device_info_bool( cl_device_id device, cl_device_info param, const char* name) {
   cl_bool a;
   clGetDeviceInfo(device, param, sizeof(cl_bool), &a, NULL);
   printf("%-40s = %s\n", name, (a?"true":"false"));
}
static void device_info_string( cl_device_id device, cl_device_info param, const char* name) {
   char a[STRING_BUFFER_LEN];
   clGetDeviceInfo(device, param, STRING_BUFFER_LEN, &a, NULL);
   printf("%-40s = %s\n", name, a);
}

// Query and display OpenCL information on device and runtime environment
static void display_device_info( cl_device_id device ) {

   printf("Querying device for info:\n");
   printf("========================\n");
   device_info_string(device, CL_DEVICE_NAME, "CL_DEVICE_NAME");
   device_info_string(device, CL_DEVICE_VENDOR, "CL_DEVICE_VENDOR");
   device_info_uint(device, CL_DEVICE_VENDOR_ID, "CL_DEVICE_VENDOR_ID");
   device_info_string(device, CL_DEVICE_VERSION, "CL_DEVICE_VERSION");
   device_info_string(device, CL_DRIVER_VERSION, "CL_DRIVER_VERSION");
   device_info_uint(device, CL_DEVICE_ADDRESS_BITS, "CL_DEVICE_ADDRESS_BITS");
   device_info_bool(device, CL_DEVICE_AVAILABLE, "CL_DEVICE_AVAILABLE");
   device_info_bool(device, CL_DEVICE_ENDIAN_LITTLE, "CL_DEVICE_ENDIAN_LITTLE");
   device_info_ulong(device, CL_DEVICE_GLOBAL_MEM_CACHE_SIZE, "CL_DEVICE_GLOBAL_MEM_CACHE_SIZE");
   device_info_ulong(device, CL_DEVICE_GLOBAL_MEM_CACHELINE_SIZE, "CL_DEVICE_GLOBAL_MEM_CACHELINE_SIZE");
   device_info_ulong(device, CL_DEVICE_GLOBAL_MEM_SIZE, "CL_DEVICE_GLOBAL_MEM_SIZE");
   device_info_bool(device, CL_DEVICE_IMAGE_SUPPORT, "CL_DEVICE_IMAGE_SUPPORT");
   device_info_ulong(device, CL_DEVICE_LOCAL_MEM_SIZE, "CL_DEVICE_LOCAL_MEM_SIZE");
   device_info_ulong(device, CL_DEVICE_MAX_CLOCK_FREQUENCY, "CL_DEVICE_MAX_CLOCK_FREQUENCY");
   device_info_ulong(device, CL_DEVICE_MAX_COMPUTE_UNITS, "CL_DEVICE_MAX_COMPUTE_UNITS");
   device_info_ulong(device, CL_DEVICE_MAX_CONSTANT_ARGS, "CL_DEVICE_MAX_CONSTANT_ARGS");
   device_info_ulong(device, CL_DEVICE_MAX_CONSTANT_BUFFER_SIZE, "CL_DEVICE_MAX_CONSTANT_BUFFER_SIZE");
   device_info_uint(device, CL_DEVICE_MAX_WORK_ITEM_DIMENSIONS, "CL_DEVICE_MAX_WORK_ITEM_DIMENSIONS");
   device_info_uint(device, CL_DEVICE_MEM_BASE_ADDR_ALIGN, "CL_DEVICE_MAX_WORK_ITEM_DIMENSIONS");
   device_info_uint(device, CL_DEVICE_MIN_DATA_TYPE_ALIGN_SIZE, "CL_DEVICE_MIN_DATA_TYPE_ALIGN_SIZE");
   device_info_uint(device, CL_DEVICE_PREFERRED_VECTOR_WIDTH_CHAR, "CL_DEVICE_PREFERRED_VECTOR_WIDTH_CHAR");
   device_info_uint(device, CL_DEVICE_PREFERRED_VECTOR_WIDTH_SHORT, "CL_DEVICE_PREFERRED_VECTOR_WIDTH_SHORT");
   device_info_uint(device, CL_DEVICE_PREFERRED_VECTOR_WIDTH_INT, "CL_DEVICE_PREFERRED_VECTOR_WIDTH_INT");
   device_info_uint(device, CL_DEVICE_PREFERRED_VECTOR_WIDTH_LONG, "CL_DEVICE_PREFERRED_VECTOR_WIDTH_LONG");
   device_info_uint(device, CL_DEVICE_PREFERRED_VECTOR_WIDTH_FLOAT, "CL_DEVICE_PREFERRED_VECTOR_WIDTH_FLOAT");
   device_info_uint(device, CL_DEVICE_PREFERRED_VECTOR_WIDTH_DOUBLE, "CL_DEVICE_PREFERRED_VECTOR_WIDTH_DOUBLE");

   {
      cl_command_queue_properties ccp;
      clGetDeviceInfo(device, CL_DEVICE_QUEUE_PROPERTIES, sizeof(cl_command_queue_properties), &ccp, NULL);
      printf("%-40s = %s\n", "Command queue out of order? ", ((ccp & CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE)?"true":"false"));
      printf("%-40s = %s\n", "Command queue profiling enabled? ", ((ccp & CL_QUEUE_PROFILING_ENABLE)?"true":"false"));
   }
}

主机程序比较长，主要执行流程为：
初始化平台、寻找设备、打印设备信息、创建设备上下文、在设备上下文中创建指令队列、载入设备代码、编译设备代码、创建核函数对象、设置核函数参数、运行核函数、等待核函数运行结束、清除所有对象。
这是OpenCL的最基本流程，虽然比较繁琐，但熟悉之后几乎每次都是这几步，代码改动很少，真正需要用心设计的是核函数。

（未完，跟帖中）

dongliudongliu

好了，再运行一个例子就睡觉。
进入上一级目录，然后切入vectorAdd，运行一下：

root@socfpga:~/helloworld# cd ..
root@socfpga:~# ls
README            helloworld        opencl_arm32_rte  vector_Add
boardtest         init_opencl.sh    swapper
root@socfpga:~# cd vector_Add/
root@socfpga:~/vector_Add# ls
vectorAdd       vectorAdd.aocx
root@socfpga:~/vector_Add# aocl program /dev/acl0 vectorAdd.aocx
aocl program: Running reprogram from /home/root/opencl_arm32_rte/board/c5soc/arm32/bin
Reprogramming was successful!
root@socfpga:~/vector_Add# ./vectorAdd
Initializing OpenCL
Platform: Altera SDK for OpenCL
Using 1 device(s)
  de1soc_sharedonly : Cyclone V SoC Development Kit
Using AOCX: vectorAdd.aocx
Launching for device 0 (1000000 elements)

Time: 107.127 ms
Kernel time (device 0): 6.933 ms

Verification: PASS



这是个向量相加的例子，也是很经典的并行计算例子。核函数内容如下：
__kernel void vectorAdd(__global const float *x,
                        __global const float *y,
                        __global float *restrict z)
{
    // get index of the work item
    int index = get_global_id(0);

    // add the vector elements
    z[index] = x[index] + y[index];
}


（未完，跟帖中）

dongliudongliu

主机程序如下：
#include <stdio.h>
#include <stdlib.h>
#include <math.h>
#include "CL/opencl.h"
#include "AOCL_Utils.h"

using namespace aocl_utils;

// OpenCL runtime configuration
cl_platform_id platform = NULL;
unsigned num_devices = 0;
scoped_array<cl_device_id> device; // num_devices elements
cl_context context = NULL;
scoped_array<cl_command_queue> queue; // num_devices elements
cl_program program = NULL;
scoped_array<cl_kernel> kernel; // num_devices elements
scoped_array<cl_mem> input_a_buf; // num_devices elements
scoped_array<cl_mem> input_b_buf; // num_devices elements
scoped_array<cl_mem> output_buf; // num_devices elements

// Problem data.
const unsigned N = 1000000; // problem size
scoped_array<scoped_aligned_ptr<float> > input_a, input_b; // num_devices elements
scoped_array<scoped_aligned_ptr<float> > output; // num_devices elements
scoped_array<scoped_array<float> > ref_output; // num_devices elements
scoped_array<unsigned> n_per_device; // num_devices elements

// Function prototypes
float rand_float();
bool init_opencl();
void init_problem();
void run();
void cleanup();

// Entry point.
int main() {
  // Initialize OpenCL.
  if(!init_opencl()) {
    return -1;
  }

  // Initialize the problem data.
  // Requires the number of devices to be known.
  init_problem();

  // Run the kernel.
  run();

  // Free the resources allocated
  cleanup();

  return 0;
}

/////// HELPER FUNCTIONS ///////

// Randomly generate a floating-point number between -10 and 10.
float rand_float() {
  return float(rand()) / float(RAND_MAX) * 20.0f - 10.0f;
}

// Initializes the OpenCL objects.
bool init_opencl() {
  cl_int status;

  printf("Initializing OpenCL\n");

  if(!setCwdToExeDir()) {
    return false;
  }

  // Get the OpenCL platform.
  platform = findPlatform("Altera");
  if(platform == NULL) {
    printf("ERROR: Unable to find Altera OpenCL platform.\n");
    return false;
  }

  // Query the available OpenCL device.
  device.reset(getDevices(platform, CL_DEVICE_TYPE_ALL, &num_devices));
  printf("Platform: %s\n", getPlatformName(platform).c_str());
  printf("Using %d device(s)\n", num_devices);
  for(unsigned i = 0; i < num_devices; ++i) {
    printf("  %s\n", getDeviceName(device[i]).c_str());
  }

  // Create the context.
  context = clCreateContext(NULL, num_devices, device, NULL, NULL, &status);
  checkError(status, "Failed to create context");

  // Create the program for all device. Use the first device as the
  // representative device (assuming all device are of the same type).
  std::string binary_file = getBoardBinaryFile("vectorAdd", device[0]);
  printf("Using AOCX: %s\n", binary_file.c_str());
  program = createProgramFromBinary(context, binary_file.c_str(), device, num_devices);

  // Build the program that was just created.
  status = clBuildProgram(program, 0, NULL, "", NULL, NULL);
  checkError(status, "Failed to build program");

  // Create per-device objects.
  queue.reset(num_devices);
  kernel.reset(num_devices);
  n_per_device.reset(num_devices);
  input_a_buf.reset(num_devices);
  input_b_buf.reset(num_devices);
  output_buf.reset(num_devices);

  for(unsigned i = 0; i < num_devices; ++i) {
    // Command queue.
    queue[i] = clCreateCommandQueue(context, device[i], CL_QUEUE_PROFILING_ENABLE, &status);
    checkError(status, "Failed to create command queue");

    // Kernel.
    const char *kernel_name = "vectorAdd";
    kernel[i] = clCreateKernel(program, kernel_name, &status);
    checkError(status, "Failed to create kernel");

    // Determine the number of elements processed by this device.
    n_per_device[i] = N / num_devices; // number of elements handled by this device

    // Spread out the remainder of the elements over the first
    // N % num_devices.
    if(i < (N % num_devices)) {
      n_per_device[i]++;
    }

    // Input buffers.
    input_a_buf[i] = clCreateBuffer(context, CL_MEM_READ_ONLY,
        n_per_device[i] * sizeof(float), NULL, &status);
    checkError(status, "Failed to create buffer for input A");

    input_b_buf[i] = clCreateBuffer(context, CL_MEM_READ_ONLY,
        n_per_device[i] * sizeof(float), NULL, &status);
    checkError(status, "Failed to create buffer for input B");

    // Output buffer.
    output_buf[i] = clCreateBuffer(context, CL_MEM_WRITE_ONLY,
        n_per_device[i] * sizeof(float), NULL, &status);
    checkError(status, "Failed to create buffer for output");
  }

  return true;
}

// Initialize the data for the problem. Requires num_devices to be known.
void init_problem() {
  if(num_devices == 0) {
    checkError(-1, "No devices");
  }

  input_a.reset(num_devices);
  input_b.reset(num_devices);
  output.reset(num_devices);
  ref_output.reset(num_devices);

  // Generate input vectors A and B and the reference output consisting
  // of a total of N elements.
  // We create separate arrays for each device so that each device has an
  // aligned buffer.
  for(unsigned i = 0; i < num_devices; ++i) {
    input_a[i].reset(n_per_device[i]);
    input_b[i].reset(n_per_device[i]);
    output[i].reset(n_per_device[i]);
    ref_output[i].reset(n_per_device[i]);

    for(unsigned j = 0; j < n_per_device[i]; ++j) {
      input_a[i][j] = rand_float();
      input_b[i][j] = rand_float();
      ref_output[i][j] = input_a[i][j] + input_b[i][j];
    }
  }
}

void run() {
  cl_int status;

  const double start_time = getCurrentTimestamp();

  // Launch the problem for each device.
  scoped_array<cl_event> kernel_event(num_devices);
  scoped_array<cl_event> finish_event(num_devices);

  for(unsigned i = 0; i < num_devices; ++i) {

    // Transfer inputs to each device. Each of the host buffers supplied to
    // clEnqueueWriteBuffer here is already aligned to ensure that DMA is used
    // for the host-to-device transfer.
    cl_event write_event[2];
    status = clEnqueueWriteBuffer(queue[i], input_a_buf[i], CL_FALSE,
        0, n_per_device[i] * sizeof(float), input_a[i], 0, NULL, &write_event[0]);
    checkError(status, "Failed to transfer input A");

    status = clEnqueueWriteBuffer(queue[i], input_b_buf[i], CL_FALSE,
        0, n_per_device[i] * sizeof(float), input_b[i], 0, NULL, &write_event[1]);
    checkError(status, "Failed to transfer input B");

    // Set kernel arguments.
    unsigned argi = 0;

    status = clSetKernelArg(kernel[i], argi++, sizeof(cl_mem), &input_a_buf[i]);
    checkError(status, "Failed to set argument %d", argi - 1);

    status = clSetKernelArg(kernel[i], argi++, sizeof(cl_mem), &input_b_buf[i]);
    checkError(status, "Failed to set argument %d", argi - 1);

    status = clSetKernelArg(kernel[i], argi++, sizeof(cl_mem), &output_buf[i]);
    checkError(status, "Failed to set argument %d", argi - 1);

    // Enqueue kernel.
    // Use a global work size corresponding to the number of elements to add
    // for this device.
    //
    // We don't specify a local work size and let the runtime choose
    // (it'll choose to use one work-group with the same size as the global
    // work-size).
    //
    // Events are used to ensure that the kernel is not launched until
    // the writes to the input buffers have completed.
    const size_t global_work_size = n_per_device[i];
    printf("Launching for device %d (%d elements)\n", i, global_work_size);

    status = clEnqueueNDRangeKernel(queue[i], kernel[i], 1, NULL,
        &global_work_size, NULL, 2, write_event, &kernel_event[i]);
    checkError(status, "Failed to launch kernel");

    // Read the result. This the final operation.
    status = clEnqueueReadBuffer(queue[i], output_buf[i], CL_FALSE,
        0, n_per_device[i] * sizeof(float), output[i], 1, &kernel_event[i], &finish_event[i]);

    // Release local events.
    clReleaseEvent(write_event[0]);
    clReleaseEvent(write_event[1]);
  }

  // Wait for all devices to finish.
  clWaitForEvents(num_devices, finish_event);

  const double end_time = getCurrentTimestamp();

  // Wall-clock time taken.
  printf("\nTime: %0.3f ms\n", (end_time - start_time) * 1e3);

  // Get kernel times using the OpenCL event profiling API.
  for(unsigned i = 0; i < num_devices; ++i) {
    cl_ulong time_ns = getStartEndTime(kernel_event[i]);
    printf("Kernel time (device %d): %0.3f ms\n", i, double(time_ns) * 1e-6);
  }

  // Release all events.
  for(unsigned i = 0; i < num_devices; ++i) {
    clReleaseEvent(kernel_event[i]);
    clReleaseEvent(finish_event[i]);
  }

  // Verify results.
  bool pass = true;
  for(unsigned i = 0; i < num_devices && pass; ++i) {
    for(unsigned j = 0; j < n_per_device[i] && pass; ++j) {
      if(fabsf(output[i][j] - ref_output[i][j]) > 1.0e-5f) {
        printf("Failed verification @ device %d, index %d\nOutput: %f\nReference: %f\n",
            i, j, output[i][j], ref_output[i][j]);
        pass = false;
      }
    }
  }

  printf("\nVerification: %s\n", pass ? "PASS" : "FAIL");
}

// Free the resources allocated during initialization
void cleanup() {
  for(unsigned i = 0; i < num_devices; ++i) {
    if(kernel && kernel[i]) {
      clReleaseKernel(kernel[i]);
    }
    if(queue && queue[i]) {
      clReleaseCommandQueue(queue[i]);
    }
    if(input_a_buf && input_a_buf[i]) {
      clReleaseMemObject(input_a_buf[i]);
    }
    if(input_b_buf && input_b_buf[i]) {
      clReleaseMemObject(input_b_buf[i]);
    }
    if(output_buf && output_buf[i]) {
      clReleaseMemObject(output_buf[i]);
    }
  }

  if(program) {
    clReleaseProgram(program);
  }
  if(context) {
    clReleaseContext(context);
  }
}

将100w维度的两个向量相加，用时107.127ms，你可以试试只用ARM计算，看需要多久，对比下性能。

好了，今天到此为止，大家晚安！

（完）