private C++ template function in a regular class - OpenCL - function

Question

I am trying to create a customized class around the OpenCL C++ wrapper to get some specific information from the available devices. For example get the number of available GPUs, CPUs, etc. in a platform. To reduce code I decided to implement a private template function as shown below:
//Devices.hpp
class Devices
{
public:
Devices(const cl::Platform& inputPlatform)
{
inputPlatform.getDevices(CL_DEVICE_TYPE_ALL, &availableDevices);
}
cl_int getTotalNumberOfDevices()
{
return availableDevices.size();
}
cl_int getTotalNumberOfGPUs()
{
return countDevicesWithSpecificProperty(CL_DEVICE_TYPE, CL_DEVICE_TYPE_GPU);
}
private:
std::vector<cl::Device> availableDevices;
template <typename T>
cl_int countDevicesWithSpecificProperty(
const cl_device_info& deviceInfo,
const T& searchPropertyValue)
{
cl_int totalNumberOfDevices = getTotalNumberOfDevices();
T response;
cl_int count = 0;
for (cl_int i = 0; i < totalNumberOfDevices; ++i)
{
try
{
availableDevices.at(i).getInfo(deviceInfo, &response);
}
catch (cl::Error e)
{
return e.err();
}
if (response == searchPropertyValue) ++count;
}
return count;
}
};
While the code compiles correctly, getInfo throws a CL_INVALID_VALUE error. When I implemented the same code using regular function (instead of a template) the code works fine:
//Devices.hpp
class Devices
{
public:
Devices(const cl::Platform& inputPlatform)
{
inputPlatform.getDevices(CL_DEVICE_TYPE_ALL, &availableDevices);
}
cl_int getTotalNumberOfDevices()
{
return availableDevices.size();
}
cl_int getTotalNumberOfGPUs()
{
return countDevicesWithSpecificProperty(CL_DEVICE_TYPE, CL_DEVICE_TYPE_GPU);
}
private:
std::vector<cl::Device> availableDevices;
cl_int countDevicesWithSpecificProperty
(const cl_device_info& deviceInfo,
const cl_device_type& searchPropertyValue)
{
cl_int totalNumberOfDevices = getTotalNumberOfDevices();
cl_device_type response;
cl_int count = 0;
for (cl_int i = 0; i < totalNumberOfDevices; ++i)
{
try
{
availableDevices.at(i).getInfo(deviceInfo, &response);
}
catch (cl::Error e)
{
return e.err();
}
if (response == searchPropertyValue) ++count;
}
return count;
}
};
Any thoughts?
PS: The method is invoked as follows:
//main.cpp
#define __CL_ENABLE_EXCEPTIONS
#include <iostream>
#include <vector>
#include <CL/cl.hpp>
#include "Devices.hpp"
int main()
{
try
{
std::vector<cl::Platform> availablePlatforms;
cl::Platform::get(&availablePlatforms);
Devices d(availablePlatforms[0]);
std::cout << d.getTotalNumberOfGPUs() << std::endl;
}
catch (cl::Error e)
{
std::cout << e.what() << std::endl << e.err() << std::endl;
}
return 0;
}

Answer 1

The issue is that your response variable doesn't have the correct type in your templated version. This is because you are passing CL_DEVICE_TYPE_GPU to the templated parameter, which is a preprocessor macro, and so won't necessarily have the correct type needed for the device info query.
One solution is to explicitly cast the templated parameter to ensure it has the correct type:
return countDevicesWithSpecificProperty(CL_DEVICE_TYPE, (cl_device_type)CL_DEVICE_TYPE_GPU);

Answer 2

Check assembler listing for both cases. There should be some difference in what compiler generates.