CODEWITHSUNDEEP
opencl
user1873073
user1873073

Reputation: 3670

Is there a good way to choose the right platform on the fly?

Because this is the computer I am using it has AMD, NVIDEA, and Intel platforms. How can I know which is the right platform to use on a users computer? What I have now is a loop that tries to create a platform, device, context, and queue for every platform. If it fails at any point it tries the next platform.

    readKernel();

    numPlatforms = getNumPlatforms(); TEST
    platforms = getPlatforms(); TEST
    for(int i = 0; i < numPlatforms; i++)
    {
        numDevices = getNumDevices(platforms[i]); TEST_AND_CONTINUE
        devices = getDevices(platforms[i], numDevices); TEST_AND_CONTINUE
        context = createContext(platforms[i], devices); TEST_AND_CONTINUE
        queue = getCommandQueue(context, devices[0]); TEST_AND_CONTINUE

        // all setup. can post info here ->  getDeviceInfo(devices[0]);
        break; 
    }
    program = createProgram(context, source); TEST
    buildProgram(program); TEST
    kernel = buildKernel(program, appName); TEST

Is that a good way to do it or is there a better way?

Upvotes: 3

Views: 2715

Answers (2)

CaptainObvious
CaptainObvious

Reputation: 2565

As usual with this kind of question, the answer is: It depends on your need. Or in other words, you need to define what is "the right platform".

Here are some cases I can think of (I'm sure anybody can find some others):

  • You developed your kernel using some features specific to a certain version of OCL. Using clGetPlatformInfo, you query each platform present to find one that has the proper OCL version.

  • You optimized your kernel for a specific type of device (CPU, GPU). You filter the devices you are interested in using the appropriate flag (CL_DEVICE_TYPE_TYPENAME) with clGetDeviceIDs.

  • You want to parallelized as much as possible the computation, but you have to move a lot of data to the device. In that case you might have found out that running your kernel on an iGPU gives the best performance. Thanks to the function clGetDeviceInfo and the flag CL_DEVICE_HOST_UNIFIED_MEMORY you can determine if you have such a device available.

  • With the clGetDeviceInfo function you can also query for a specific vendor extension that you want to use (flag: CL_DEVICE_EXTENSIONS). Note that clGetPlatformInfo provides also a list of extension supported by the platform.

  • You have several GPUs available and you want the one with the "best performance". Still with clGetDeviceInfo you can query certain specifications of the device. Based on these specs you can make you choice. For instance you can found out if the device has cache (CL_DEVICE_GLOBAL_MEM_CACHE_TYPE) and if yes how much (CL_DEVICE_GLOBAL_MEM_CACHELINE_SIZE). You can also query the clock frequency (CL_DEVICE_MAX_CLOCK_FREQUENCY) or how many compute units are available on the device (CL_DEVICE_MAX_COMPUTE_UNITS).

Upvotes: 8

DarkZeros
DarkZeros

Reputation: 8410

Typically a good common use case is to:

  1. Get all the platforms
  2. Get each platform GPU and CPU/Other devices, separate them into 2 arrays.
  3. Is there a GPU device available? Select that platform&device
  4. Is there a CPU/Other device available? Select that platform&device

You can refine the 3 and 4 points, to selec only the best GPU device depending on your needs with clGetDeviceInfo().

Upvotes: 4

Related Questions