gpuinfo reports wrong workitem dimensions for 32-bit architectures

[doe300]

The value returned by clGetDeviceInfo with the parameter CL_DEVICE_MAX_WORK_ITEM_SIZES is of type size_t[] and is read by easycl::printDeviceInfoArray() as cl_ulong[].
Since for 32-bit systems, size_t is unsigned int, not unsigned long, the value read is wrong.

[hughperkins]

Ok. Thoughts on how to fix this?

[doe300]

If all device-info queries which use easycl::printDeviceInfoArray() have the type size_t[], you could just rewrite this method to use a size_t array.

Otherwise, copy easycl::printDeviceInfoArray() and write a version for arrays of type size_t. Or make it a template-function with specializations for both cl_ulongand size_t and use the correct one.

[hughperkins]

First option sounds tempting, but it's not very general, is that a fair impression?

Second option sounds plausible, and easy-ish. Would need to keep the old function in, for backwards compatibility.

One variant on the second is, looks like clGetDeviceInfo can tell us the size in bytes of the data being returned? https://www.khronos.org/registry/OpenCL/sdk/1.0/docs/man/xhtml/clGetDeviceInfo.html

[doe300]

Yeah, 1 is the more simple one, but will introduce other errors, if for some query cl_ulong[] is returned. Which, as far as I can tell, is never used as a return-type for a device-query in OpenCL 2.1, but may be used by extensions or future versions.

As to getting the byte-size of the type returned: The type-size can be calculated, and then you could use formatted printing (with either %uor %ul) depending on the byte-size of the type.

[hughperkins]

As to getting the byte-size of the type returned: The type-size can be calculated, and then you could use formatted printing (with either %uor %ul) depending on the byte-size of the type.

Yes. Well, at least have a case switch, with the size_t and cl_long types for now, and an exception for other types.