OpenCL设置nvidia gpu

    用OpenCV去处理图像的时候，为了提高效率会使用UMat去处理。

    UMat会调用OpenCL。但由于系统环境设置原因，可能一直会调用集成显卡。

    查看ocl.cpp源码后发现有下列方法。其中主要关注getenv，调用系统环境。

static cl_device_id selectOpenCLDevice()
{
    std::string platform, deviceName;
    std::vector deviceTypes;

    const char* configuration = getenv("OPENCV_OPENCL_DEVICE");
    if (configuration &&
            (strcmp(configuration, "disabled") == 0 ||
             !parseOpenCLDeviceConfiguration(std::string(configuration), platform, deviceTypes, deviceName)
            ))
        return NULL;

    bool isID = false;
    int deviceID = -1;
    if (deviceName.length() == 1)
    // We limit ID range to 0..9, because we want to write:
    // - '2500' to mean i5-2500
    // - '8350' to mean AMD FX-8350
    // - '650' to mean GeForce 650
    // To extend ID range change condition to '> 0'
    {
        isID = true;
        for (size_t i = 0; i < deviceName.length(); i++)
        {
            if (!isdigit(deviceName[i]))
            {
                isID = false;
                break;
            }
        }
        if (isID)
        {
            deviceID = atoi(deviceName.c_str());
            if (deviceID < 0)
                return NULL;
        }
    }

    std::vector platforms;
    {
        cl_uint numPlatforms = 0;
        CV_OCL_DBG_CHECK(clGetPlatformIDs(0, NULL, &numPlatforms));

        if (numPlatforms == 0)
            return NULL;
        platforms.resize((size_t)numPlatforms);
        CV_OCL_DBG_CHECK(clGetPlatformIDs(numPlatforms, &platforms[0], &numPlatforms));
        platforms.resize(numPlatforms);
    }

    int selectedPlatform = -1;
    if (platform.length() > 0)
    {
        for (size_t i = 0; i < platforms.size(); i++)
        {
            std::string name;
            CV_OCL_DBG_CHECK(getStringInfo(clGetPlatformInfo, platforms[i], CL_PLATFORM_NAME, name));
            if (name.find(platform) != std::string::npos)
            {
                selectedPlatform = (int)i;
                break;
            }
        }
        if (selectedPlatform == -1)
        {
            std::cerr << "ERROR: Can't find OpenCL platform by name: " << platform << std::endl;
            goto not_found;
        }
    }
    if (deviceTypes.size() == 0)
    {
        if (!isID)
        {
            deviceTypes.push_back("GPU");
            if (configuration)
                deviceTypes.push_back("CPU");
        }
        else
            deviceTypes.push_back("ALL");
    }
    for (size_t t = 0; t < deviceTypes.size(); t++)
    {
        int deviceType = 0;
        std::string tempStrDeviceType = deviceTypes[t];
        std::transform(tempStrDeviceType.begin(), tempStrDeviceType.end(), tempStrDeviceType.begin(), char_tolower);

        if (tempStrDeviceType == "gpu" || tempStrDeviceType == "dgpu" || tempStrDeviceType == "igpu")
            deviceType = Device::TYPE_GPU;
        else if (tempStrDeviceType == "cpu")
            deviceType = Device::TYPE_CPU;
        else if (tempStrDeviceType == "accelerator")
            deviceType = Device::TYPE_ACCELERATOR;
        else if (tempStrDeviceType == "all")
            deviceType = Device::TYPE_ALL;
        else
        {
            std::cerr << "ERROR: Unsupported device type for OpenCL device (GPU, CPU, ACCELERATOR): " << deviceTypes[t] << std::endl;
            goto not_found;
        }

        std::vector devices; // TODO Use clReleaseDevice to cleanup
        for (int i = selectedPlatform >= 0 ? selectedPlatform : 0;
                (selectedPlatform >= 0 ? i == selectedPlatform : true) && (i < (int)platforms.size());
                i++)
        {
            cl_uint count = 0;
            cl_int status = clGetDeviceIDs(platforms[i], deviceType, 0, NULL, &count);
            if (!(status == CL_SUCCESS || status == CL_DEVICE_NOT_FOUND))
            {
                CV_OCL_DBG_CHECK_RESULT(status, "clGetDeviceIDs get count");
            }
            if (count == 0)
                continue;
            size_t base = devices.size();
            devices.resize(base + count);
            status = clGetDeviceIDs(platforms[i], deviceType, count, &devices[base], &count);
            if (!(status == CL_SUCCESS || status == CL_DEVICE_NOT_FOUND))
            {
                CV_OCL_DBG_CHECK_RESULT(status, "clGetDeviceIDs get IDs");
            }
        }

        for (size_t i = (isID ? deviceID : 0);
             (isID ? (i == (size_t)deviceID) : true) && (i < devices.size());
             i++)
        {
            std::string name;
            CV_OCL_DBG_CHECK(getStringInfo(clGetDeviceInfo, devices[i], CL_DEVICE_NAME, name));
            cl_bool useGPU = true;
            if(tempStrDeviceType == "dgpu" || tempStrDeviceType == "igpu")
            {
                cl_bool isIGPU = CL_FALSE;
                CV_OCL_DBG_CHECK(clGetDeviceInfo(devices[i], CL_DEVICE_HOST_UNIFIED_MEMORY, sizeof(isIGPU), &isIGPU, NULL));
                useGPU = tempStrDeviceType == "dgpu" ? !isIGPU : isIGPU;
            }
            if ( (isID || name.find(deviceName) != std::string::npos) && useGPU)
            {
                // TODO check for OpenCL 1.1
                return devices[i];
            }
        }
    }

not_found:
    if (!configuration)
        return NULL; // suppress messages on stderr

    std::cerr << "ERROR: Requested OpenCL device not found, check configuration: " << configuration << std::endl
            << "    Platform: " << (platform.length() == 0 ? "any" : platform) << std::endl
            << "    Device types: ";
    for (size_t t = 0; t < deviceTypes.size(); t++)
        std::cerr << deviceTypes[t] << " ";

    std::cerr << std::endl << "    Device name: " << (deviceName.length() == 0 ? "any" : deviceName) << std::endl;
    return NULL;
}

    如果想要调用nvidia gpu， 最容易的方法，只需如下配置即可：

set OPENCV_OPENCL_DEVICE=NVIDIA:GPU:0

具体含义可以查看以下代码：

// Layout: ::
// Sample: AMD:GPU:
// Sample: AMD:GPU:Tahiti
// Sample: :GPU|CPU: = '' = ':' = '::'
static bool parseOpenCLDeviceConfiguration(const std::string& configurationStr,
        std::string& platform, std::vector& deviceTypes, std::string& deviceNameOrID)
{
    std::vector parts;
    split(configurationStr, ':', parts);
    if (parts.size() > 3)
    {
        std::cerr << "ERROR: Invalid configuration string for OpenCL device" << std::endl;
        return false;
    }
    if (parts.size() > 2)
        deviceNameOrID = parts[2];
    if (parts.size() > 1)
    {
        split(parts[1], '|', deviceTypes);
    }
    if (parts.size() > 0)
    {
        platform = parts[0];
    }
    return true;
}