document error for kernel enqueue with a work-group size of zero

[bashbaug]

possible fix for #1541

[lpavank]

Looks good to me.

[karolherbst]

I wonder if requiring this could break applications that simply specify a [0, 0, 0] local size.

But the spec also states If local_work_size is specified, the values specified in global_work_size[0], …​, global_work_size[work_dim - 1] must be evenly divisible by the corresponding values specified in local_work_size[0], …​, local_work_size[work_dim - 1] and that would be a division by 0. So if we have some major vendor that don't handle explicit local sizes of 0 I think it's safe to merge.

However if those are handled it's difficult to rule out that it wouldn't break things. But I think this is good to clarify, I just know that in rusticl enqueue with local_work_size == [0, 0, 0] is the same as specifying NULL. Not because the intention was to support 0 local sizes, but because it made sense to implement it that way logistically.

[bashbaug]

So if we have some major vendor that don't handle explicit local sizes of 0 I think it's safe to merge.

Mentioning for completeness: we currently return the error CL_INVALID_WORK_GROUP_SIZE when the local work-group size is explicitly zero for both CPU and GPU.

>>>> clEnqueueNDRangeKernel( Test ): queue = 0x5aefdf259830, kernel = 0x5aefe19b8810, global_work_size = < 512 x 1 >, local_work_size = < 0 x 0 >
<<<< clEnqueueNDRangeKernel -> CL_INVALID_WORK_GROUP_SIZE