很久以前写过有关D3D11和OCL直接共享显存的代码，
Intel MediaSDK sample_decode 官方GPU解码流程学习 - DirectX11 D3D11和OpenCL共享资源
这段代码一直运行的很好，被我用来做验证显卡驱动里的D3D11和OCL兼容性的一个测试项目。
最近发现这个测试程序在Intel双显卡的笔记本上运行不正常，当命令行指定使用集显UHD630或者独显Intel Iris(R) Xe MAX Graphics上做硬件解码的时候
sample_decode h264 -i test.h264 -hw -d3d11 -r -iGfx[-dGfx]总会有一块显卡在运行时会失败，调试了一下代码，发现是在初始化OpenCL的时候，在调用clGetDeviceIDsFromD3D11KHR()时返回-1 (CL_DEVICE_NOT_FOUND),因此后面创建OCL的clContext和clCommandQueue之类的操作就都失败了。
status = ptrToFunction_clGetDeviceIDsFromD3D11KHR(g_platformToUse, CL_D3D11_DEVICE_KHR, (void *)g_pd3dDevice, CL_PREFERRED_DEVICES_FOR_D3D11_KHR, 0, NULL, &numDevs);
printf("clGetDeviceIDsFromD3D11KHR returns status = %d\n", status);
运行输出
clGetDeviceIDsFromD3D11KHR returns status = -1clGetDeviceIDsFromD3D11KHR - Querying OpenCL Devices Corresponding to Direct3D 11 Devices.
cl_int clGetDeviceIDsFromD3D11KHR(cl_platform_id platform,
cl_d3d11_device_source_khr d3d_device_source,
void *d3d_object,
cl_d3d11_device_set_khr d3d_device_set,
cl_uint num_entries, cl_device_id *devices,
cl_uint *num_devices);
ERRORS
Returns CL_SUCCESS if the function is executed successfully. Otherwise it may return:
· CL_INVALID_PLATFORM if platform is not a valid platform.
· CL_INVALID_VALUE if d3d_device_source is not a valid value, d3d_device_set is not a
valid value, num_entries is equal to zero and devices is not NULL, or if both
num_devices and devices are NULL.
· CL_DEVICE_NOT_FOUND if no OpenCL devices that correspond to d3d_object were found.
返回CL_DEVICE_NOT_FOUND说明在传进去的参数 d3d_object 和 platform 之间不匹配。
跟踪了一下代码， d3d_object是通过D3D11CreateDevice()创建的，由传进去的-iGfx/-dGfx控制创建在独显或者集显硬件设备上
platform这个参数就比较有趣了，现在的代码流程是这样
//[1] get the platform
g_platformToUse = NULL;
status = clGetPlatformIDs(0, NULL, &numPlatforms);
testStatus(status, "clGetPlatformIDs error");
cl_platform_id *platforms = (cl_platform_id *)malloc(sizeof(cl_platform_id) * numPlatforms);
if(platforms == NULL)
{
printf("Error when allocating space for the platforms\n");
exit(EXIT_FAILURE);
}
status = clGetPlatformIDs(numPlatforms, platforms, NULL);
testStatus(status, "clGetPlatformIDs error");
//[2] get device ids for the platform i have obtained
cl_uint num_devices = -1; //yeah yeah, i know uint
cl_device_info devTypeToUse = CL_DEVICE_CPU_OR_GPU; //set as CPU or GPU from host_common.h
for(unsigned int i=0;i<numPlatforms;i++)
{
char pbuf[256];
status = clGetPlatformInfo(platforms[i], CL_PLATFORM_VENDOR, sizeof(pbuf), pbuf, NULL);
testStatus(status, "clGetPlatformInfo error");
//leaving this strcmp for AMD, used when debugging and is the precise AMD platform name
//if(!strcmp(pbuf, "Advanced Micro Devices, Inc."))
//We found the Intel Platform, this is the one we wantto use
if(!strcmp(pbuf, "Intel(R) Corporation"))
{
printf("Great! We found an Intel OpenCL platform.\n");
g_platformToUse = platforms[i];
status = clGetDeviceIDs(g_platformToUse, devTypeToUse, 0, g_clDevices, &num_devices);
if (status == CL_DEVICE_NOT_FOUND)
continue;
else
break;
}
}对于单显卡或者只有1块intel显卡的时候，这个流程是没问题的。但是在有2块Intel显卡的机器上，枚举ocl platformd的数量会返回2；而且2块ocl platform的vendor名字都包含"Intel(R) Corporation"，但是根据这个逻辑，代码永远只会用搜到的第一个包含"Intel"字串的设备作为后面的platform 参数。（而且我发现根据安装驱动的方法和流程不同，这个默认的第一个OCL platform有时候是UHD630，有时候是Iris Xe MAX...)。如果d3d11设备是创建在UHD630上的，而opencl platform是对应的Iris Xe MAX, 调用clGetDeviceIDsFromD3D11KHR()就一定会返回-1, 这也很好理解，2个框架创建在不同的硬件设备上，显存当然不能共享访问了。
要解决这个问题，只能在OCL框架里把每个platform下面的device打印出来，看看怎么和d3d11设备匹配... 这段代码要改
//获取每个platform下面的设备数量
status = clGetDeviceIDs(platforms[i], CL_DEVICE_TYPE_ALL, num_devices, clDevices, 0);
...
for(cl_uint iDevNum=0;iDevNum<num_devices;iDevNum++)
{
//获取每个设备的一些OCL属性
clGetDeviceInfo(clDevices[iDevNum], CL_DEVICE_TYPE, sizeof(cl_device_type), (void *)&device_type, NULL);
clGetDeviceInfo(clDevices[iDevNum], CL_DEVICE_VENDOR, (sizeof(char)*256), &vendorName, NULL);
clGetDeviceInfo(clDevices[iDevNum], CL_DEVICE_EXTENSIONS, (sizeof(char)*2048), &devExtString, NULL);
...
...
...
}接着发现了一个问题，OpenCL里面只能获取CL_DEVICE_VENDOR_ID, 就是0x8086, 获取不了device id.
D3D11 API这边可以获取vendor id和device id
DXGI_ADAPTER_DESC adapterDesc;
adapter->GetDesc(&adapterDesc);
if (wcsstr(adapterDesc.Description, L"Intel"))
{
intel = adapterIdx;
printf("Found %d: %ls VendorId:%08X DeviceId:%08X\n", adapterIdx,adapterDesc.Description, adapterDesc.VendorId, adapterDesc.DeviceId);
}
运行输出:
Found 0: Intel(R) UHD Graphics 630 VendorId:00008086 DeviceId:00003E98为了找到OpenCL下面获取device id的方法，再度和谷哥讨论了一下，发现OpenCL标准里确实没有这个功能，但是对于双显卡的场景，这又是个刚需。所以A家N家都是通过OCL Extension的方式增加了一个接口Associating OpenCL device ids with GPUs |Anteru's Blog
Intel显卡这边同样也是增加了一个接口，在OpenCL 3.0的Extension里 cl_intel_device_attribute_query
所以用之前需要查询OpenCL是否支持cl_intel_device_attribute_query这个extension. 这个需要驱动支持
status = clGetDeviceIDs(platforms[i], CL_DEVICE_TYPE_ALL, num_devices, clDevices, 0);
testStatus(status, "clGetDeviceIDs error");
for(cl_uint iDevNum=0;iDevNum<num_devices;iDevNum++)
{
//query each for their extension string
//print out the device type just to make sure we got it right
cl_device_type device_type;
char vendorName[256];
memset(vendorName, '\0', 256);
char devExtString[2048];
memset(devExtString, '\0', 2048);
clGetDeviceInfo(clDevices[iDevNum], CL_DEVICE_TYPE, sizeof(cl_device_type), (void *)&device_type, NULL);
clGetDeviceInfo(clDevices[iDevNum], CL_DEVICE_VENDOR, (sizeof(char)*256), &vendorName, NULL);
clGetDeviceInfo(clDevices[iDevNum], CL_DEVICE_EXTENSIONS, (sizeof(char)*2048), &devExtString, NULL);
//查询当前显卡驱动是否支持cl_khr_d3d11_sharing扩展
char *extStringStart1 = NULL;
extStringStart1 = strstr(devExtString, "cl_khr_d3d11_sharing");
...
//查询当前显卡驱动是否支持cl_intel_device_attribute_query扩展
char *extStringStart2 = NULL;
extStringStart2 = strstr(devExtString, "cl_intel_device_attribute_query");
if (extStringStart2 != 0)
{
if (device_type == CL_DEVICE_TYPE_CPU)
{
strcpy_s(devTypeString, "CPU");
}
else if (device_type == CL_DEVICE_TYPE_GPU)
{
strcpy_s(devTypeString, "GPU");
}
else
{
strcpy_s(devTypeString, "Not a CPU or GPU"); //for sample code, not product
}
printf("Device %s in %s platform supports cl_intel_device_attribute_query extension\n", devTypeString, vendorName);
}
else
{
printf("Device %s in %s platform does not support cl_intel_device_attribute_query extension\n", devTypeString, vendorName);
}
} //end for(...)如果支持这个extension, 接下来就可以获取device id了
//print out the device type just to make sure we got it right
cl_device_type device_type;
cl_uint device_vendor_id;
cl_uint device_device_id;
char vendorName[256];
char DeviceVersion[1024];
memset(vendorName, '\0', 256);
memset(DeviceVersion, '\0', 1024);
#define CL_DEVICE_ID_INTEL 0x4251
clGetDeviceInfo(dummy_g_clDevices[dummy_device_id], CL_DEVICE_TYPE, sizeof(cl_device_type), (void *)&device_type, NULL);
clGetDeviceInfo(dummy_g_clDevices[dummy_device_id], CL_DEVICE_VENDOR_ID, sizeof(cl_uint), (void *)&device_vendor_id, NULL);
clGetDeviceInfo(dummy_g_clDevices[dummy_device_id], CL_DEVICE_ID_INTEL, sizeof(cl_uint), (void *)&device_device_id, NULL);
clGetDeviceInfo(dummy_g_clDevices[dummy_device_id], CL_DEVICE_VENDOR, (sizeof(char) * 256), vendorName, NULL);
clGetDeviceInfo(dummy_g_clDevices[dummy_device_id], CL_DEVICE_VERSION, (sizeof(char) * 1024), DeviceVersion, NULL);
if (device_type == CL_DEVICE_TYPE_CPU)
{
printf(" [%d] Device type is CPU, Vendor is %s\n", dummy_device_id, vendorName);
}
else if (device_type == CL_DEVICE_TYPE_GPU)
{
printf(" [%d] Device type is GPU, Vendor is %s\n", dummy_device_id, vendorName);
}
else
{
printf(" [%d] device type is unknown\n",dummy_device_id );
}
clGetDeviceInfo(dummy_g_clDevices[0], CL_DEVICE_NAME, (sizeof(char) * 256), vendorName, NULL);
printf(" [%d] Device name is %s\n", dummy_device_id, vendorName);
printf(" [%d] Device_vendor_id is 0x%08X\n", dummy_device_id, device_vendor_id);
printf(" [%d] device_device_id is 0x%08X\n", dummy_device_id, device_device_id);
printf(" [%d] CL_DEVICE_VERSION is %s\n", dummy_device_id, DeviceVersion);
运行输出:
Great! We found an Intel OpenCL platform. - platform index[0]
platform index[0] : num_devices = 1
[0] Device type is GPU, Vendor is Intel(R) Corporation
[0] Device name is Intel(R) UHD Graphics 630
[0] Device_vendor_id is 0x00008086
[0] device_device_id is 0x00003E98
[0] CL_DEVICE_VERSION is OpenCL 3.0 NEO
Great! We found an Intel OpenCL platform. - platform index[1]
platform index[1] : num_devices = 1
[0] Device type is GPU, Vendor is Intel(R) Corporation
[0] Device name is Intel(R) Iris(R) Xe MAX Graphics
[0] Device_vendor_id is 0x00008086
[0] device_device_id is 0x00004905
[0] CL_DEVICE_VERSION is OpenCL 3.0 NEO从2个platform 里分别获取了2个显卡的device id, 有了这些信息，我们就知道该把哪个platform作为参数传给clGetDeviceIDsFromD3D11KHR()了
上面这种模式算是基本解决了问题，但是也发现了一个新的问题：这个OCL获取device id的扩展是在opencl 3.0的标准里加进去的，也就是说只支持OpenCL 2.0的老的显卡驱动里，如果查询显卡属性不支持"cl_intel_device_attribute_query"的时候是无法获取到这个device id的。比如下面这个输出是在一个非常老的显卡驱动上跑的，此时驱动只支持到opencl 2.1。
Great! We found an Intel OpenCL platform. - platform index[0]
platform index[0] : num_devices = 1
[0] Device type is GPU, Vendor is Intel(R) Corporation
[0] Device name is Intel(R) UHD Graphics 630
[0] Device_vendor_id is 0x00008086
[0] device_device_id is 0x00000000
[0] CL_DEVICE_VERSION is OpenCL 2.1 NEO对于这种情况，如果要求每一台机器都升级到最新的显卡驱动，显然是不现实的。所以就想到了另一个方法，既然一般家用电脑最多也就是双显卡的配置，那么前面枚举OpenCL platform的时候，最多也就会获得2个platform, 那么索性就把这2个platform依次传给clGetDeviceIDsFromD3D11KHR()匹配，如果不匹配就会返回CL_DEVICE_NOT_FOUND, 如果返回CL_SUCCESS就继续进行下面的OpenCL的初始化。
//获取当前平台下面的cl platform数量和信息
std::vector < cl_platform_id> intel_platforms;
intel_platforms.empty();
//[1] get the platform
g_platformToUse = NULL;
status = clGetPlatformIDs(0, NULL, &numPlatforms);
testStatus(status, "clGetPlatformIDs error");
cl_platform_id *platforms = (cl_platform_id *)malloc(sizeof(cl_platform_id) * numPlatforms);
if(platforms == NULL)
{
printf("Error when allocating space for the platforms\n");
exit(EXIT_FAILURE);
}
status = clGetPlatformIDs(numPlatforms, platforms, NULL);
testStatus(status, "clGetPlatformIDs error");
printf("Total numPlatforms found = %d\n", numPlatforms);
for(unsigned int i=0;i<numPlatforms;i++)
{
char pbuf[256];
//获取这个platform的vendro描述字符串
status = clGetPlatformInfo(platforms[i], CL_PLATFORM_VENDOR, sizeof(pbuf), pbuf, NULL);
testStatus(status, "clGetPlatformInfo error");
...
if(!strcmp(pbuf, "Intel(R) Corporation"))
{
//找到一个intel platform
printf("Great! We found an Intel OpenCL platform. - platform index[%d]\n",i);
...
//将符合条件的platform加入到intel_platforms这个vector容器里
intel_platforms.push_back(platforms[i]);
...
}
}
...
...
...
cl_int platform_index;
//用遍历的方式依次匹配d3d11设备和opencl platform
for (cl_platform_id intel_platform : intel_platforms)
{
g_platformToUse = intel_platform;
cps[1] = (cl_context_properties)g_platformToUse;
ptrToFunction_clGetDeviceIDsFromD3D11KHR = NULL;
ptrToFunction_clGetDeviceIDsFromD3D11KHR = (clGetDeviceIDsFromD3D11KHR_fn)clGetExtensionFunctionAddressForPlatform(g_platformToUse, "clGetDeviceIDsFromD3D11KHR");
numDevs = 0;
//size_t bytes = 0;
//careful with the g_pd3DDevice
status = ptrToFunction_clGetDeviceIDsFromD3D11KHR(g_platformToUse, CL_D3D11_DEVICE_KHR, (void *)g_pd3dDevice, CL_PREFERRED_DEVICES_FOR_D3D11_KHR, 0, NULL, &numDevs);
printf("clGetDeviceIDsFromD3D11KHR returns status = %d\n", status);
std::vector <cl_platform_id> ::iterator itr;
itr = find(intel_platforms.begin(), intel_platforms.end(), intel_platform);
cl_int platform_index = distance(intel_platforms.begin(), itr);
if (status != CL_SUCCESS)
{
//不匹配，继续匹配下一个platform
printf("Failed: the [%d/%d] OCL platform unmatch\n", platform_index + 1, intel_platforms.size());
}
else
{
//匹配成功,跳出循环，进行后面的初始化
printf("Good: Found the matched OCL platform [%d/%d]\n", platform_index + 1, intel_platforms.size());
break;
}
}
//检查是否匹配到正确的cl platform
testStatus(status, "Failed on clGetDeviceIDsFromD3D11KHR");最后测试一下遍历匹配的代码
没啥大问题，搞定收工 ^_^
最后还是老规矩，源码附上，仅供参考 :)