模型推理OpenCL 使用流程解析
Posted 极智视界
tags:
篇首语：本文由小常识网(cha138.com)小编为大家整理，主要介绍了模型推理OpenCL 使用流程解析相关的知识，希望对你有一定的参考价值。
欢迎关注我的公众号 [极智视界]，获取我的更多笔记分享
大家好，我是极智视界。本文介绍了 OpenCL 的使用流程。
下面开始。
文章目录
1、整体流程
2、具体流程
cpp main 函数：
/*
主函数
*/
int main(int argc,char **argv)
cl_context context = 0;
cl_command_queue commandQueue = 0;
cl_program program = 0;
cl_device_id device = 0;
cl_kernel kernel = 0;
cl_mem memObjects[3] = 0,0,0 ;
cl_int errNum;
// 创建OpenCL上下文
context = CreateContext(&device);
if (context == NULL)
printf("Failed to create OpenCL context." );
return 1;
// 获得OpenCL设备,并创建命令队列
commandQueue = CreateCommandQueue(context, device);
if (commandQueue == NULL)
Cleanup(context,commandQueue,program, kernel,
memObjects);
return 1;
// 创建OPenCL程序
program = CreateProgram(context,device, "vecAdd.cl");
if (program == NULL)
Cleanup(context,commandQueue,program,
kernel,memObjects);
return 1;
// 创建OpenCL内核
kernel = clCreateKernel(program, "vector_add",NULL);
if (kernel == NULL)
printf( "Failed to create kernel");
Cleanup(context,commandQueue,program,
kernel,memObjects);
return 1;
// 创建OpenCL内存对象
float result[ARRAY_SIZE];
float a[ARRAY_SIZE];
float b[ARRAY_SIZE];
for (int i = 0; i < ARRAY_SIZE; i++)
a[i] = (float)i;
b[i] = (float)(i * 2);
if (!CreateMemObjects(context,memObjects, a,b))
Cleanup(context,commandQueue,program,
kernel,memObjects);
return 1;
// 设置内核参数
errNum = clSetKernelArg(kernel,0,sizeof (cl_mem),
&memObjects[0]);
errNum |= clSetKernelArg(kernel,1,sizeof (cl_mem),
&memObjects[1]);
errNum |= clSetKernelArg(kernel,2,sizeof (cl_mem),
&memObjects[2]);
if (errNum!= CL_SUCCESS)
printf("Error setting kernel arguments.");
Cleanup(context,commandQueue,program,
kernel,memObjects);
return 1;
size_t globalWorkSize[1] = ARRAY_SIZE ;
size_t localWorkSize[1] = 1 ;
// 执行内核
errNum = clEnqueueNDRangeKernel (commandQueue,kernel,
1,NULL,
globalWorkSize,
localWorkSize,0,NULL,
NULL);
if (errNum!= CL_SUCCESS)
printf( "Error queuing kernel for execution." );
Cleanup(context,commandQueue,program, kernel,
memObjects);
return 1;
// 计算结果拷贝回主机
errNum = clEnqueueReadBuffer(commandQueue, memObjects[2],
CL_TRUE,0,
ARRAY_SIZE * sizeof(float),
result,0, NULL,NULL);
if (errNum!= CL_SUCCESS)
printf( "Error reading result buffer." );
Cleanup(context,commandQueue,program, kernel,
memObjects);
return 1;
for (int i = 0; i < ARRAY_SIZE; i++)
printf("i=%d：%f\\n",i,result[i]);
printf("Executed program succesfully." );
Cleanup(context,commandQueue,program, kernel,memObjects);
return 0;
下面是 .cl 文件中的 opencl 核函数，以实现向量加为例：
__kernel
void vector_add(global const float *a,global const float *b,global float *result)
int gid = get_global_id(0);
result[gid] = a[gid] + b[gid];
接下来是分过程的各个阶段的主要函数。
1) 寻找设备，创建上下文
/*
1.创建平台
2.创建设备
3.根据设备创建上下文
*/
cl_context CreateContext(cl_device_id *device)
cl_int errNum;
cl_uint numPlatforms;
cl_platform_id firstPlatformId;
cl_context context = NULL;
errNum = clGetPlatformIDs(1, &firstPlatformId,&numPlatforms);
if (errNum!= CL_SUCCESS || numPlatforms <= 0)
printf( "Failed to find any OpenCL platforms." );
return NULL;
errNum = clGetDeviceIDs(firstPlatformId, CL_DEVICE_TYPE_GPU,1,
device,NULL);
if (errNum!= CL_SUCCESS)
printf( "There is no GPU,trying CPU……" );
errNum = clGetDeviceIDs(firstPlatformId,
CL_DEVICE_TYPE_CPU, 1,device,NULL);
if (errNum!= CL_SUCCESS)
printf( "There is NO GPU or CPU" );
return NULL;
context = clCreateContext(NULL,1, device, NULL,NULL,& errNum);
if (errNum!= CL_SUCCESS)
printf( " create context error\\n" );
return NULL;
return context;
2) 创建队列命令
/*
@在上下文可用的第一个设备中创建命令队列
*/
cl_command_queue CreateCommandQueue(cl_context context,
cl_device_id device)
cl_int errNum;
cl_command_queue commandQueue = NULL;
commandQueue = clCreateCommandQueue(context, device,0,NULL);
if (commandQueue == NULL)
printf("Failed to create commandQueue for device 0");
return NULL;
return commandQueue;
#define _FILE_ "vecAdd.cl"
#define _LINE_ 20
const int ARRAY_SIZE = 10;
char *ReadKernelSourceFile(const char *filename, size_t *length)
FILE *file = NULL;
size_t sourceLength;
char *sourceString;
int ret;
file = fopen(filename,"rb");
if(file == NULL)
printf("%s at %d：Can't open %s\\n",_FILE_, _LINE_ - 2,
filename);
return NULL;
fseek(file,0,SEEK_END);
sourceLength = ftell(file);
fseek(file,0,SEEK_SET);
sourceString = (char *)malloc(sourceLength + 1);
sourceString[0] = '\\0';
ret = fread(sourceString,sourceLength,1, file);
if(ret == 0)
printf("%s at %d：Can't read source %s\\n", _FILE_,
_LINE_ - 2,filename);
return NULL;
fclose(file);
if(length!= 0)
*length = sourceLength;
sourceString[sourceLength] = '\\0';
return sourceString;
/*
@读取内核源码创建OpenCL程序
*/
cl_program CreateProgram(cl_context context,
cl_device_id device,
const char *fileName)
cl_int errNum;
cl_program program;
size_t program_length;
//从.cl文件中获取cl代码
char *const source = ReadKernelSourceFile(fileName,
&program_length);
// 创建程序对象
program = clCreateProgramWithSource(context, 1,
(const char **)&source,
NULL, NULL);
if (program == NULL)
printf("Failed to create CL program from source." );
return NULL;
// 编译程序对象
errNum = clBuildProgram(program,0,NULL, NULL,NULL,NULL);
if (errNum!= CL_SUCCESS)
char buildLog[16384];
clGetProgramBuildInfo(program,device,
CL_PROGRAM_BUILD_LOG,
sizeof(buildLog),
buildLog,NULL);
printf("Error in kernel：%s ",buildLog);
clReleaseProgram(program);
return NULL;
return program;
3) 创建存储器对象
/*
@创建内存对象
*/
bool CreateMemObjects(cl_context context,cl_mem memObjects[3],
float * a, float * b)
// clCreateBuffer会在上下文中创建存储器对象，也就是内存对象，而是否设置主机内存、全局、全局常量、局部、私有内存等，通过设置cl_mem_flags(eg:CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR)
memObjects[0] = clCreateBuffer(context,
CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR,
sizeof(float) * ARRAY_SIZE,a,NULL);
memObjects[1] = clCreateBuffer(context,
CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR,
sizeof(float) * ARRAY_SIZE,b,NULL);
memObjects[2] = clCreateBuffer(context, CL_MEM_READ_WRITE,
sizeof(float) * ARRAY_SIZE,
NULL,NULL);
if (memObjects[0] == NULL || memObjects[1] == NULL
|| memObjects[2] == NULL)
printf("Error creating memory objects." );
return false;
return true;
4) 释放资源函数
/*
@清除OpenCL资源
*/
void Cleanup(cl_context context, cl_command_queue commandQueue,
cl_program program,cl_kernel kernel,
cl_mem memObjects[3])
for (int i = 0; i < 3; i++)
if (memObjects[i]!= 0)
clReleaseMemObject(memObjects[i]);
if (commandQueue!= 0)
clReleaseCommandQueue(commandQueue);
if (kernel!= 0)
clReleaseKernel(kernel);
if (program!= 0)
clReleaseProgram(program);
if (context!= 0)
clReleaseContext(context);
好了，以上分享了 OpenCL 基本使用流程，希望我的分享能对你的学习有一点帮助。
【公众号传送】
扫描下方二维码即可关注我的微信公众号【极智视界】，获取更多AI经验分享，让我们用极致+极客的心态来迎接AI ！
以上是关于模型推理OpenCL 使用流程解析的主要内容，如果未能解决你的问题，请参考以下文章
我的NVIDIA开发者之旅 - 极智AI | TensorRT API 构建模型推理流程
模型推理寒武纪 MLU resnet50 量化及离线推理流程