如何使用 OPENCL 在多 GPU 环境中唯一标识一个 GPU？
How to uniquely identify a GPU in a multiGPU environment using OPENCL?
我正在学习 OpenCL 并且已经编写了初步代码来查询机器并找出与之关联的平台和设备。
#include <stdio.h>
#include <stdlib.h>
#include <CL/cl.h>
int main(int argc,char** argv){
printf("Let's see what devices are there in this Node\n\n");
cl_int errNum,errCPU,errGPU;
cl_uint numPlatforms;
cl_platform_id *platformIds;
cl_context context=NULL;
char dname[500];
int i,dc,dg;
cl_device_id *cpuDevices,*gpuDevices;
cl_uint numCPUDevices,numGPUDevices,entries;
cl_ulong long_entries;
size_t p_size;
errNum = clGetPlatformIDs(0,NULL,&numPlatforms);
if(errNum==CL_SUCCESS){ printf("Number of Platforms on this Node: %d\n\n",numPlatforms); }
else{ printf("Error:Failure in clGetPlatformIDs,error code=%d\n",errNum); }
platformIds = (cl_platform_id *)malloc(sizeof(cl_platform_id)*numPlatforms);
errNum = clGetPlatformIDs(numPlatforms,platformIds,NULL);
if(errNum==CL_SUCCESS){
for(i=0;i<numPlatforms;i++){
printf("Platform Information on %d Platform\n",i+1);
/*Obtain information about platform*/
clGetPlatformInfo(platformIds[i],CL_PLATFORM_NAME,500,dname,NULL);
printf("\tCL_PLATFORM_NAME = %s\n",dname);
clGetPlatformInfo(platformIds[i],CL_PLATFORM_VERSION,500,dname,NULL);
printf("\tCL_PLATFORM_VERSION = %s\n",dname);
/*obtain list of devices available on platform*/
clGetDeviceIDs(platformIds[i],CL_DEVICE_TYPE_CPU,0,NULL,&numCPUDevices);
printf("\t%d CPU devices found\n",numCPUDevices);
clGetDeviceIDs(platformIds[i],CL_DEVICE_TYPE_GPU,0,NULL,&numGPUDevices);
printf("\t%d GPU devices found\n",numGPUDevices);
cpuDevices = (cl_device_id *)malloc(sizeof(cl_device_id)*numCPUDevices);
gpuDevices = (cl_device_id *)malloc(sizeof(cl_device_id)*numGPUDevices);
printf("\tPrinting devices Information\n");
if(numCPUDevices>0){
for(dc=0;dc<numCPUDevices;dc++){
printf("\t\tPrinting CPU Devices Information\n");
errCPU = clGetDeviceIDs(platformIds[i],CL_DEVICE_TYPE_CPU,1,&cpuDevices[dc],NULL);
if(errCPU==CL_SUCCESS){
printf("\t\tDevice Id is %d\n",cpuDevices[dc]);
printf("\t\tDevice Information of %d device on %d platform\n",dc+1,i+1);
clGetDeviceInfo(cpuDevices[dc],CL_DEVICE_NAME,500,dname,NULL);
printf("\t\tDevice # %d name = %s\n",dc,dname);
clGetDeviceInfo(cpuDevices[dc],CL_DRIVER_VERSION,500,dname,NULL);
printf("\t\tDriver version = %s\n",dname);
clGetDeviceInfo(cpuDevices[dc],CL_DEVICE_GLOBAL_MEM_SIZE,sizeof(cl_ulong),&long_entries,NULL);
printf("\t\tGlobal Memory (MB) : \t%llu\n",long_entries/1024/1024);
clGetDeviceInfo(cpuDevices[dc],CL_DEVICE_GLOBAL_MEM_CACHE_SIZE,sizeof(cl_ulong),&long_entries,NULL);
printf("\t\tGlobal Memory Cache (MB):\t%llu\n",long_entries/1024/1024);
clGetDeviceInfo(cpuDevices[dc],CL_DEVICE_LOCAL_MEM_SIZE,sizeof(cl_ulong),&long_entries,NULL);
printf("\t\tLocal Memory (KB): \t%llu\n",long_entries/1024);
clGetDeviceInfo(cpuDevices[dc],CL_DEVICE_MAX_CLOCK_FREQUENCY,sizeof(cl_ulong),&long_entries,NULL);
printf("\t\tMax clock (MHz) : \t%llu\n",long_entries);
clGetDeviceInfo(cpuDevices[dc],CL_DEVICE_MAX_WORK_GROUP_SIZE,sizeof(size_t),&p_size,NULL);
printf("\t\tMax Work Group Size: \t%d\n",p_size);
clGetDeviceInfo(cpuDevices[dc],CL_DEVICE_MAX_COMPUTE_UNITS,sizeof(cl_uint),&entries,NULL);
printf("\t\tNumber of parallel compute cores:\t%d\n",entries);
}else{printf("\t\tError:Failure in clGetDeviceIds,error code = %d\n",errCPU);}
}
}else{printf("\t\tZero CPU Devices found\n");}
/* query devices for information */
if(numGPUDevices>0){
for(dg = 0;dg<numGPUDevices;dg++ ){
printf("\t\tPrinting GPU Devices Information\n");
errGPU = clGetDeviceIDs(platformIds[i],CL_DEVICE_TYPE_GPU,1,&gpuDevices[dg],NULL);
if(errGPU==CL_SUCCESS){
printf("\t\tDevice Id is %d\n",gpuDevices[dg]);
printf("\t\tDevice Information of %d device on %d platform\n",dg+1,i+1);
clGetDeviceInfo(gpuDevices[dg],CL_DEVICE_NAME,500,dname,NULL);
printf("\t\tDevice # %d name = %s\n",dg,dname);
clGetDeviceInfo(gpuDevices[dg],CL_DRIVER_VERSION,500,dname,NULL);
printf("\t\tDriver version = %s\n",dname);
clGetDeviceInfo(gpuDevices[dg],CL_DEVICE_GLOBAL_MEM_SIZE,sizeof(cl_ulong),&long_entries,NULL);
printf("\t\tGlobal Memory (MB) : \t%llu\n",long_entries/1024/1024);
clGetDeviceInfo(gpuDevices[dg],CL_DEVICE_GLOBAL_MEM_CACHE_SIZE,sizeof(cl_ulong),&long_entries,NULL);
printf("\t\tGlobal Memory Cache (MB):\t%llu\n",long_entries/1024/1024);
clGetDeviceInfo(gpuDevices[dg],CL_DEVICE_LOCAL_MEM_SIZE,sizeof(cl_ulong),&long_entries,NULL);
printf("\t\tLocal Memory (KB): \t%llu\n",long_entries/1024);
clGetDeviceInfo(gpuDevices[dg],CL_DEVICE_MAX_CLOCK_FREQUENCY,sizeof(cl_ulong),&long_entries,NULL);
printf("\t\tMax clock (MHz) : \t%llu\n",long_entries);
clGetDeviceInfo(gpuDevices[dg],CL_DEVICE_MAX_WORK_GROUP_SIZE,sizeof(size_t),&p_size,NULL);
printf("\t\tMax Work Group Size: \t%d\n",p_size);
clGetDeviceInfo(gpuDevices[dg],CL_DEVICE_MAX_COMPUTE_UNITS,sizeof(cl_uint),&entries,NULL);
printf("\t\tNumber of parallel compute cores:\t%d\n",entries);
}else{printf("\t\tError:Failure in clGetDeviceIds,error code = %d\n\t\t,Or,This platform cannot interact with the GPUs.Check for the drivers\n",errGPU);}
}
}else{printf("\t\tZero GPU Devices found\n");}
}
}
else{
printf("Error:Failure in clGetPlatformIDs,error code = %d\n",errNum);
}
return 0;
}
当我执行它时，我得到了这样的示例输出
Let's see what devices are there in this Node
Number of Platforms on this Node: 2
Platform Information on 1 Platform
CL_PLATFORM_NAME = AMD Accelerated Parallel Processing
CL_PLATFORM_VERSION = OpenCL 2.1 AMD-APP (2527.3)
0 CPU devices found
3 GPU devices found
Printing devices Information
Zero CPU Devices found
Printing GPU Devices Information
Device Id is 13401120
Device Information of 1 device on 1 platform
Device # 0 name = gfx900
Driver version = 2527.3 (HSA1.1,HSAIL)
Global Memory (MB) : 16368
Global Memory Cache (MB): 0
Local Memory (KB): 64
Max clock (MHz) : 1500
Max Work Group Size: 256
Number of parallel compute cores: 64
Printing GPU Devices Information
Device Id is 13401120
Device Information of 2 device on 1 platform
Device # 1 name = gfx900
Driver version = 2527.3 (HSA1.1,HSAIL)
Global Memory (MB) : 16368
Global Memory Cache (MB): 0
Local Memory (KB): 64
Max clock (MHz) : 1500
Max Work Group Size: 256
Number of parallel compute cores: 64
Printing GPU Devices Information
Device Id is 13401120
Device Information of 3 device on 1 platform
Device # 2 name = gfx900
Driver version = 2527.3 (HSA1.1,HSAIL)
Global Memory (MB) : 16368
Global Memory Cache (MB): 0
Local Memory (KB): 64
Max clock (MHz) : 1500
Max Work Group Size: 256
Number of parallel compute cores: 64
Platform Information on 2 Platform
CL_PLATFORM_NAME = Intel(R) OpenCL
CL_PLATFORM_VERSION = OpenCL 2.0 LINUX
1 CPU devices found
3 GPU devices found
Printing devices Information
Printing CPU Devices Information
Device Id is 16035224
Device Information of 1 device on 2 platform
Device # 0 name = Intel(R) Xeon(R) Platinum 8164 CPU @ 2.00GHz
Driver version = 1.2.0.37
Global Memory (MB) : 47782
Global Memory Cache (MB): 0
Local Memory (KB): 32
Max clock (MHz) : 2000
Max Work Group Size: 8192
Number of parallel compute cores: 52
Printing GPU Devices Information
Error:Failure in clGetDeviceIds,error code = -1
,Or,This platform cannot interact with the GPUs.Check for the drivers
Printing GPU Devices Information
Error:Failure in clGetDeviceIds,error code = -1
,Or,This platform cannot interact with the GPUs.Check for the drivers
Printing GPU Devices Information
Error:Failure in clGetDeviceIds,error code = -1
,Or,This platform cannot interact with the GPUs.Check for the drivers
我的问题，为什么三个gpu的deviceID都一样。如果我输入 %p，我会得到十六进制的答案。
一般来说，如何在给定的平台上唯一标识一个设备，以便我可以 运行 内核在那个特定的设备上？假设，我想在多 GPU 环境中将具有特定数据的内核发送到一个 GPU？
for (dg in gpuDevices)
errGPU = clGetDeviceIDs(platformIds[i],CL_DEVICE_TYPE_GPU,1,&gpuDevices[dg],NULL);
这是在请求单个设备 ID（始终返回第一个）并将其写入 gpuDevices
中的不同位置
将clGetDeviceIDs从设备循环中拉出来，取而代之的是运行它在获取设备数量后立即执行。这将立即填充所有设备 ID。
gpuDevices = (cl_device_id *)malloc(sizeof(cl_device_id)*numGPUDevices);
errGPU = clGetDeviceIDs(platformIds[i], CL_DEVICE_TYPE_GPU, numGPUDevices, gpuDevices, NULL)
我正在学习 OpenCL 并且已经编写了初步代码来查询机器并找出与之关联的平台和设备。
#include <stdio.h>
#include <stdlib.h>
#include <CL/cl.h>
int main(int argc,char** argv){
printf("Let's see what devices are there in this Node\n\n");
cl_int errNum,errCPU,errGPU;
cl_uint numPlatforms;
cl_platform_id *platformIds;
cl_context context=NULL;
char dname[500];
int i,dc,dg;
cl_device_id *cpuDevices,*gpuDevices;
cl_uint numCPUDevices,numGPUDevices,entries;
cl_ulong long_entries;
size_t p_size;
errNum = clGetPlatformIDs(0,NULL,&numPlatforms);
if(errNum==CL_SUCCESS){ printf("Number of Platforms on this Node: %d\n\n",numPlatforms); }
else{ printf("Error:Failure in clGetPlatformIDs,error code=%d\n",errNum); }
platformIds = (cl_platform_id *)malloc(sizeof(cl_platform_id)*numPlatforms);
errNum = clGetPlatformIDs(numPlatforms,platformIds,NULL);
if(errNum==CL_SUCCESS){
for(i=0;i<numPlatforms;i++){
printf("Platform Information on %d Platform\n",i+1);
/*Obtain information about platform*/
clGetPlatformInfo(platformIds[i],CL_PLATFORM_NAME,500,dname,NULL);
printf("\tCL_PLATFORM_NAME = %s\n",dname);
clGetPlatformInfo(platformIds[i],CL_PLATFORM_VERSION,500,dname,NULL);
printf("\tCL_PLATFORM_VERSION = %s\n",dname);
/*obtain list of devices available on platform*/
clGetDeviceIDs(platformIds[i],CL_DEVICE_TYPE_CPU,0,NULL,&numCPUDevices);
printf("\t%d CPU devices found\n",numCPUDevices);
clGetDeviceIDs(platformIds[i],CL_DEVICE_TYPE_GPU,0,NULL,&numGPUDevices);
printf("\t%d GPU devices found\n",numGPUDevices);
cpuDevices = (cl_device_id *)malloc(sizeof(cl_device_id)*numCPUDevices);
gpuDevices = (cl_device_id *)malloc(sizeof(cl_device_id)*numGPUDevices);
printf("\tPrinting devices Information\n");
if(numCPUDevices>0){
for(dc=0;dc<numCPUDevices;dc++){
printf("\t\tPrinting CPU Devices Information\n");
errCPU = clGetDeviceIDs(platformIds[i],CL_DEVICE_TYPE_CPU,1,&cpuDevices[dc],NULL);
if(errCPU==CL_SUCCESS){
printf("\t\tDevice Id is %d\n",cpuDevices[dc]);
printf("\t\tDevice Information of %d device on %d platform\n",dc+1,i+1);
clGetDeviceInfo(cpuDevices[dc],CL_DEVICE_NAME,500,dname,NULL);
printf("\t\tDevice # %d name = %s\n",dc,dname);
clGetDeviceInfo(cpuDevices[dc],CL_DRIVER_VERSION,500,dname,NULL);
printf("\t\tDriver version = %s\n",dname);
clGetDeviceInfo(cpuDevices[dc],CL_DEVICE_GLOBAL_MEM_SIZE,sizeof(cl_ulong),&long_entries,NULL);
printf("\t\tGlobal Memory (MB) : \t%llu\n",long_entries/1024/1024);
clGetDeviceInfo(cpuDevices[dc],CL_DEVICE_GLOBAL_MEM_CACHE_SIZE,sizeof(cl_ulong),&long_entries,NULL);
printf("\t\tGlobal Memory Cache (MB):\t%llu\n",long_entries/1024/1024);
clGetDeviceInfo(cpuDevices[dc],CL_DEVICE_LOCAL_MEM_SIZE,sizeof(cl_ulong),&long_entries,NULL);
printf("\t\tLocal Memory (KB): \t%llu\n",long_entries/1024);
clGetDeviceInfo(cpuDevices[dc],CL_DEVICE_MAX_CLOCK_FREQUENCY,sizeof(cl_ulong),&long_entries,NULL);
printf("\t\tMax clock (MHz) : \t%llu\n",long_entries);
clGetDeviceInfo(cpuDevices[dc],CL_DEVICE_MAX_WORK_GROUP_SIZE,sizeof(size_t),&p_size,NULL);
printf("\t\tMax Work Group Size: \t%d\n",p_size);
clGetDeviceInfo(cpuDevices[dc],CL_DEVICE_MAX_COMPUTE_UNITS,sizeof(cl_uint),&entries,NULL);
printf("\t\tNumber of parallel compute cores:\t%d\n",entries);
}else{printf("\t\tError:Failure in clGetDeviceIds,error code = %d\n",errCPU);}
}
}else{printf("\t\tZero CPU Devices found\n");}
/* query devices for information */
if(numGPUDevices>0){
for(dg = 0;dg<numGPUDevices;dg++ ){
printf("\t\tPrinting GPU Devices Information\n");
errGPU = clGetDeviceIDs(platformIds[i],CL_DEVICE_TYPE_GPU,1,&gpuDevices[dg],NULL);
if(errGPU==CL_SUCCESS){
printf("\t\tDevice Id is %d\n",gpuDevices[dg]);
printf("\t\tDevice Information of %d device on %d platform\n",dg+1,i+1);
clGetDeviceInfo(gpuDevices[dg],CL_DEVICE_NAME,500,dname,NULL);
printf("\t\tDevice # %d name = %s\n",dg,dname);
clGetDeviceInfo(gpuDevices[dg],CL_DRIVER_VERSION,500,dname,NULL);
printf("\t\tDriver version = %s\n",dname);
clGetDeviceInfo(gpuDevices[dg],CL_DEVICE_GLOBAL_MEM_SIZE,sizeof(cl_ulong),&long_entries,NULL);
printf("\t\tGlobal Memory (MB) : \t%llu\n",long_entries/1024/1024);
clGetDeviceInfo(gpuDevices[dg],CL_DEVICE_GLOBAL_MEM_CACHE_SIZE,sizeof(cl_ulong),&long_entries,NULL);
printf("\t\tGlobal Memory Cache (MB):\t%llu\n",long_entries/1024/1024);
clGetDeviceInfo(gpuDevices[dg],CL_DEVICE_LOCAL_MEM_SIZE,sizeof(cl_ulong),&long_entries,NULL);
printf("\t\tLocal Memory (KB): \t%llu\n",long_entries/1024);
clGetDeviceInfo(gpuDevices[dg],CL_DEVICE_MAX_CLOCK_FREQUENCY,sizeof(cl_ulong),&long_entries,NULL);
printf("\t\tMax clock (MHz) : \t%llu\n",long_entries);
clGetDeviceInfo(gpuDevices[dg],CL_DEVICE_MAX_WORK_GROUP_SIZE,sizeof(size_t),&p_size,NULL);
printf("\t\tMax Work Group Size: \t%d\n",p_size);
clGetDeviceInfo(gpuDevices[dg],CL_DEVICE_MAX_COMPUTE_UNITS,sizeof(cl_uint),&entries,NULL);
printf("\t\tNumber of parallel compute cores:\t%d\n",entries);
}else{printf("\t\tError:Failure in clGetDeviceIds,error code = %d\n\t\t,Or,This platform cannot interact with the GPUs.Check for the drivers\n",errGPU);}
}
}else{printf("\t\tZero GPU Devices found\n");}
}
}
else{
printf("Error:Failure in clGetPlatformIDs,error code = %d\n",errNum);
}
return 0;
}
当我执行它时，我得到了这样的示例输出
Let's see what devices are there in this Node
Number of Platforms on this Node: 2
Platform Information on 1 Platform
CL_PLATFORM_NAME = AMD Accelerated Parallel Processing
CL_PLATFORM_VERSION = OpenCL 2.1 AMD-APP (2527.3)
0 CPU devices found
3 GPU devices found
Printing devices Information
Zero CPU Devices found
Printing GPU Devices Information
Device Id is 13401120
Device Information of 1 device on 1 platform
Device # 0 name = gfx900
Driver version = 2527.3 (HSA1.1,HSAIL)
Global Memory (MB) : 16368
Global Memory Cache (MB): 0
Local Memory (KB): 64
Max clock (MHz) : 1500
Max Work Group Size: 256
Number of parallel compute cores: 64
Printing GPU Devices Information
Device Id is 13401120
Device Information of 2 device on 1 platform
Device # 1 name = gfx900
Driver version = 2527.3 (HSA1.1,HSAIL)
Global Memory (MB) : 16368
Global Memory Cache (MB): 0
Local Memory (KB): 64
Max clock (MHz) : 1500
Max Work Group Size: 256
Number of parallel compute cores: 64
Printing GPU Devices Information
Device Id is 13401120
Device Information of 3 device on 1 platform
Device # 2 name = gfx900
Driver version = 2527.3 (HSA1.1,HSAIL)
Global Memory (MB) : 16368
Global Memory Cache (MB): 0
Local Memory (KB): 64
Max clock (MHz) : 1500
Max Work Group Size: 256
Number of parallel compute cores: 64
Platform Information on 2 Platform
CL_PLATFORM_NAME = Intel(R) OpenCL
CL_PLATFORM_VERSION = OpenCL 2.0 LINUX
1 CPU devices found
3 GPU devices found
Printing devices Information
Printing CPU Devices Information
Device Id is 16035224
Device Information of 1 device on 2 platform
Device # 0 name = Intel(R) Xeon(R) Platinum 8164 CPU @ 2.00GHz
Driver version = 1.2.0.37
Global Memory (MB) : 47782
Global Memory Cache (MB): 0
Local Memory (KB): 32
Max clock (MHz) : 2000
Max Work Group Size: 8192
Number of parallel compute cores: 52
Printing GPU Devices Information
Error:Failure in clGetDeviceIds,error code = -1
,Or,This platform cannot interact with the GPUs.Check for the drivers
Printing GPU Devices Information
Error:Failure in clGetDeviceIds,error code = -1
,Or,This platform cannot interact with the GPUs.Check for the drivers
Printing GPU Devices Information
Error:Failure in clGetDeviceIds,error code = -1
,Or,This platform cannot interact with the GPUs.Check for the drivers
我的问题，为什么三个gpu的deviceID都一样。如果我输入 %p，我会得到十六进制的答案。
一般来说，如何在给定的平台上唯一标识一个设备，以便我可以 运行 内核在那个特定的设备上？假设，我想在多 GPU 环境中将具有特定数据的内核发送到一个 GPU？
for (dg in gpuDevices) errGPU = clGetDeviceIDs(platformIds[i],CL_DEVICE_TYPE_GPU,1,&gpuDevices[dg],NULL);
这是在请求单个设备 ID（始终返回第一个）并将其写入 gpuDevices
将clGetDeviceIDs从设备循环中拉出来，取而代之的是运行它在获取设备数量后立即执行。这将立即填充所有设备 ID。
gpuDevices = (cl_device_id *)malloc(sizeof(cl_device_id)*numGPUDevices);
errGPU = clGetDeviceIDs(platformIds[i], CL_DEVICE_TYPE_GPU, numGPUDevices, gpuDevices, NULL)