OpenCLC程序提供bizzare输出
我已经写了一个简单的OpenCLC代码。其核心代码是:OpenCLC程序提供bizzare输出,c,linux,parallel-processing,opencl,gpgpu,C,Linux,Parallel Processing,Opencl,Gpgpu,我已经写了一个简单的OpenCLC代码。其核心代码是: __kernel void hello(__global int * A,__global int * B) { int x=get_global_id(0); B[x]=x; A[x]+=1; } 以下是主机代码的一部分: int main() { cl_platform_id* platforms=NULL; cl_device_id* devices=NULL; cl_uint ret
__kernel void hello(__global int * A,__global int * B)
{
int x=get_global_id(0);
B[x]=x;
A[x]+=1;
}
以下是主机代码的一部分:
int main()
{
cl_platform_id* platforms=NULL;
cl_device_id* devices=NULL;
cl_uint ret,platformCount,deviceCount;
cl_context context = NULL;
cl_command_queue command_queue=NULL;
char* name;
int i,j,l;
size_t size;
cl_mem memobj = NULL;
cl_mem memobj1 = NULL;
cl_program program = NULL;
cl_kernel kernel = NULL;
int array_size=10;
int *A=(int*)malloc(array_size*sizeof(int));
int *B=(int*)malloc(array_size*sizeof(int));
printf("\nhey");
for(i=0;i<array_size;i++)
A[i]=0;
ret=clGetPlatformIDs(0,NULL,&platformCount);
printf("\n-----------------Found %d platforms-----------\n",platformCount);
platforms=(cl_platform_id*)malloc(sizeof(cl_platform_id)*platformCount);
ret=clGetPlatformIDs(platformCount,platforms,NULL);
if(ret==CL_SUCCESS)
printf("\nPlatform ids obtained successfully!");
for(i=platformCount-1;i>=0;i--)
{
char * platformname;
printf("\n-------------In platform %d-----------------",i);
ret=clGetPlatformInfo(platforms[i], CL_PLATFORM_NAME ,0,NULL,&size);
platformname=(char*)malloc(sizeof(char)*size);
ret=clGetPlatformInfo(platforms[i], CL_PLATFORM_NAME ,size,platformname,NULL);
if(ret==CL_SUCCESS)
printf("\nPlatform info obtained successfully!");
printf("\n----------------For %s-----------------------",platformname);
ret=clGetDeviceIDs(platforms[i],CL_DEVICE_TYPE_DEFAULT,NULL,NULL,&deviceCount);
devices=(cl_device_id*)malloc(sizeof(cl_device_id)*deviceCount);
ret=clGetDeviceIDs(platforms[i],CL_DEVICE_TYPE_DEFAULT,deviceCount,devices,NULL);
if(ret==CL_SUCCESS)
printf("\nFound %d devices!",deviceCount);
for(j=0;j<deviceCount;j++)
{
char*devicename;
printf("\n-------------Device %d.%d-----------------",i,j+1);
ret=clGetDeviceInfo(devices[j], CL_DEVICE_NAME ,0,NULL,&size);
devicename=(char*)malloc(sizeof(char)*size);
ret=clGetDeviceInfo(devices[j], CL_DEVICE_NAME ,size,devicename,NULL);
if(ret==CL_SUCCESS)
printf("\nDevice info obtained successfully!");
printf("\nThe device name is %s && size=%d\n",devicename,size);
printf("\nFound %d corresponding devices",deviceCount);
context=clCreateContext(NULL,1,&devices[j], NULL, NULL, &ret);
if(ret==CL_SUCCESS)
printf("\nContext created successfully");
command_queue=clCreateCommandQueue(context,devices[j], CL_QUEUE_PROFILING_ENABLE,&ret);
if(ret==CL_SUCCESS)
printf("\nCommand queue created successfully");
memobj=clCreateBuffer(context, CL_MEM_READ_WRITE ,array_size*sizeof(int),NULL,&ret);
if(ret==CL_SUCCESS)
printf("\nMemory object 1 created successfully");
memobj1=clCreateBuffer(context, CL_MEM_READ_WRITE ,array_size*sizeof(int),NULL,&ret);
if(ret==CL_SUCCESS)
printf("\nMemory object 2 created successfully");
ret=clEnqueueWriteBuffer(command_queue, memobj, CL_TRUE, 0, array_size*sizeof(int), A, 0, NULL, NULL);
if(ret==CL_SUCCESS)
printf("\nData written into buffer1 successfully");
ret=clEnqueueWriteBuffer(command_queue, memobj1, CL_TRUE, 0, array_size*sizeof(int),B, 0, NULL, NULL);
if(ret==CL_SUCCESS)
printf("\nData written into buffer2 successfully");
FILE *fp;
char fileName[] = "./6.cl";
char *source_str;
size_t source_size;
/* Load kernel code */
fp = fopen(fileName, "r");
if (!fp) {
fprintf(stderr, "Failed to load kernel.\n");
exit(1);
}
source_str = (char*)malloc(MAX_SOURCE_SIZE);
source_size = fread(source_str, 1, MAX_SOURCE_SIZE, fp);
printf("\nThe program is \n%s\n",source_str);
fclose(fp);
program=clCreateProgramWithSource(context, 1, (const char **)&source_str,(const size_t *)&source_size, &ret);
if(ret==CL_SUCCESS)
printf("\nProgram created successfully");
ret = clBuildProgram(program, 1, &devices[j], NULL, NULL, NULL);
if(ret==CL_SUCCESS)
printf("\nProgram built successfully");
ret=clGetProgramBuildInfo(program,devices[j], CL_PROGRAM_BUILD_STATUS ,0,NULL,&size);
printf("\n Program buildinfo status=%d",ret);
cl_build_status *status=(cl_build_status *)malloc(sizeof(cl_build_status )*size);
clGetProgramBuildInfo(program,devices[j], CL_PROGRAM_BUILD_STATUS ,size,status,NULL);
printf("\nBuild status=%d\n",*status);
printf("\nBuild log i=%d, j=%d",i,j);
ret=clGetProgramBuildInfo(program,devices[j], CL_PROGRAM_BUILD_LOG ,0,NULL,&size);
printf("\nclGetProgramBuildInfo ret1=%d",ret);
char buildlog[2048];
ret=clGetProgramBuildInfo(program,devices[j], CL_PROGRAM_BUILD_LOG ,sizeof(buildlog),buildlog,NULL);
printf("\nclGetProgramBuildInfo ret2=%d",ret);
printf("\n!!!!!!!!!!!!!!!!!!!!!Program ended!!!!!!!!!!!\n");
printf("\n\nBuildlog: %s\n\n",buildlog);
kernel = clCreateKernel(program, "hello", &ret);
if(ret==CL_SUCCESS)
printf("\nKernel created successfully");
ret = clSetKernelArg(kernel, 0, sizeof(cl_mem), (void *) &memobj);
printf("\nKernel argument 1=%d",ret);
ret = clSetKernelArg(kernel, 1, sizeof(cl_mem), (void *) &memobj1);
printf("\nKernel argument 2=%d",ret);
cl_uint work_dim = 1;
size_t global_item_size=array_size;
size_t local_item_size=32;
cl_event event;
ret = clEnqueueNDRangeKernel(command_queue, kernel, work_dim, NULL,&global_item_size, &local_item_size,0,NULL,&event);
if(ret==CL_SUCCESS)
printf("\nKernel executed successfully");
//ret=clEnqueueTask(command_queue,kernel,0,NULL,NULL);
clWaitForEvents(1, &event);//make sure kernel has finished
clFinish(command_queue);//make sure all enqueued tasks finished
//get the profiling data and calculate the kernel execution time
cl_ulong time_start, time_end;
double total_time;
clGetEventProfilingInfo(event, CL_PROFILING_COMMAND_START, sizeof(time_start), &time_start, NULL);
clGetEventProfilingInfo(event, CL_PROFILING_COMMAND_END, sizeof(time_end), &time_end, NULL);
total_time = (cl_double)(time_end - time_start)/1000000.0;
printf("OpenCl Execution time is: %10.5f[ms] \n",total_time);
ret = clEnqueueReadBuffer(command_queue, memobj1, CL_TRUE, 0,array_size * sizeof(int), B, 0, NULL, NULL);
ret = clEnqueueReadBuffer(command_queue, memobj, CL_TRUE, 0,array_size * sizeof(int), A, 0, NULL, NULL);
int t;
printf("\nThe result is:");
for(t=0;t<array_size;t++)
printf("\t%d",B[t]);
printf("\nThe result A is:");
for(t=0;t<array_size;t++)
printf("\t%d",A[t]);
//Read file here
}
}
return 0;
}
intmain()
{
cl_platform_id*platforms=NULL;
cl_设备_id*设备=NULL;
cl_uint ret、平台计数、设备计数;
cl_上下文=空;
cl_命令_队列命令_队列=NULL;
字符*名称;
int i,j,l;
大小;
cl_mem memobj=NULL;
cl_mem memobj1=NULL;
cl_程序=空;
cl_内核=NULL;
int数组_size=10;
int*A=(int*)malloc(数组大小*sizeof(int));
int*B=(int*)malloc(数组大小*sizeof(int));
printf(“\nhey”);
对于(i=0;i=0;i--)
{
char*平台名;
printf(“\n-----------------在平台%d------------中”,i);
ret=clGetPlatformInfo(平台[i],CL_平台名称,0,NULL和大小);
platformname=(char*)malloc(sizeof(char)*size);
ret=clGetPlatformInfo(平台[i],平台名称,大小,平台名称,空);
如果(ret==CL_成功)
printf(“\n成功获取平台信息!”);
printf(“\n--------------对于%s-----------------”,platformname);
ret=CLGetDeviceID(平台[i]、CL\U设备类型\U默认值、NULL、NULL和deviceCount);
设备=(cl_设备id*)malloc(sizeof(cl_设备id)*设备计数);
ret=CLGetDeviceID(平台[i],CL_设备类型默认,设备计数,设备,空);
如果(ret==CL_成功)
printf(“\n找到%d个设备!”,设备计数);
对于(j=0;j
您正在传递一个空(NULL
)事件等待列表,但声称其中有1个事件。如果您正在检查OpenCL API调用中的错误代码(您应该一直这样做),则此操作将失败,您很可能会返回CL\u INVALID\u EVENT\u WAIT\u LIST
,这将为您指出问题所在
如果您只想从内核检索事件(例如用于分析),但不想向其传递任何事件依赖项,则正确的形式如下:
ret = clEnqueueNDRangeKernel(command_queue, kernel, work_dim, NULL,&global_item_size, &local_item_size,0,NULL,&event);
您提供的代码的第二个问题是,您的工作组大小不能精确划分全局大小:
size_t global_item_size=array_size; // which is 10
size_t local_item_size=32;
这将导致clEnqueueNDRangeKernel
返回CL\u无效的工作组大小
,并且无法将任何工作排队
代码列出的特定平台的设备比实际可用的设备要少
您正在请求类型为CL\u DEVICE\u type\u DEFAULT
的所有设备。如果您确实需要所有设备,请使用CL\u DEVICE\u type\u all
。如果您只需要GPU设备,请请求类型为CL\u DEVICE\u type\u GPU这是可行的。它在两个GPU上都提供了正确的输出,但在我运行此co时会提供垃圾值CPU上的de。您是否检查每个OpenCL API调用返回的错误代码?如果您检查了,并且它们都是CL_SUCCESS
,请使用完整的主机代码更新您的问题,我可以再看一看。更新:此代码在第一次运行时给出垃圾值,而与使用的设备无关。实际上,我正在运行两个loops viz 1 for(i=platformCount-1;i>=0;i--){/*获取阵列设备中相应的设备ID*/for(j=0;jOK),检查所有错误代码,如果它们都很好,那么用一个完整的、最小的示例来更新问题,重现您的问题。@ikk用第二个问题更新了答案。修复后,您的代码对我来说很好。
size_t global_item_size=array_size; // which is 10
size_t local_item_size=32;