[petsc-dev] OpenCL platform and device query routines

[Karl Rupp]

Hi Barry,

thanks, I will use -opencl_show_platforms in the meanwhile then.
-opencl_show_devices would make sense as an alias for 
-opencl_show_platforms only - we can always add that later if required.

Best regards,
Karli


On 04/12/2013 10:35 PM, Barry Smith wrote:
>
>    We currently have
>
> -cuda_show_devices
> -cuda_set_device
>
> in init.c
>
>    why not, for now, do something in analogy for OpenCL?
>
>     We can decide later if "view" is the appropriate keyword to use?
>
>     Barry
>
>
> On Apr 12, 2013, at 4:51 PM, Karl Rupp <rupp at mcs.anl.gov> wrote:
>
>> Dear PETScians,
>>
>> in order to make proper use of OpenCL functionality, we need some diagnostics for the user such that the correct device is used. Such functionality is partly also desired for CUDA, but less urgent (only addresses NVIDIA GPUs anyway).
>>
>> OpenCL defines platforms (think of it as SDKs from the various vendors) and devices with one type out of {CPU, GPU, ACCELERATOR}. Each platform may support multiple devices, but not necessarily all OpenCL-enabled devices on the machine. For example, the AMD SDK (platform) does not provide support for NVIDIA GPUs, but it supports Intel CPUs (x86 ftw!). Since multiple SDKs can be installed in parallel, information on the proper enumeration is quite important to use the correct device.
>>
>> Example: A machine equipped with an Intel CPU and an NVIDIA GPU with OpenCL SDKs from Intel, AMD, and NVIDIA installed. Within OpenCL one will 'see' the following:
>>
>> - Platform 0:
>>   - Vendor: Intel
>>   - Device 0: Intel i7 whatever (CPU)
>>
>> - Platform 1:
>>   - Vendor: AMD
>>   - Device 0: Intel i7 whatever (CPU)
>>
>> - Platform 2:
>>   - Vendor NVIDIA
>>   - Device 0: NVIDIA GTX whatever (GPU)
>>
>> (Maybe in different order. Matters can get worse with Xeon Phi, AMD APUs, etc.)
>>
>> To provide the necessary diagnostics, I suggest in line with -vec_view the flag
>>   -opencl_view
>> to print the OpenCL infrastructure available on the system.
>> Is there any better naming scheme/proposal? -cuda_view and maybe some time later -threadcomm_view (-numa_view?) would follow from this choice. Note that this should be independent of external linear algebra libraries such as CUSP, ViennaCL, etc. to avoid unnecessary code duplication. However, the actual platform/device *setter* flags (e.g. pick device 0 from platform 1) need to be package-specific.
>>
>> Best regards,
>> Karli
>