[SYCL][L0][CUDA][HIP] Fix PI_KERNEL_GROUP_INFO_GLOBAL_WORK_SIZE queries

[abagusetty]

Address kernel query global_work_size for L0, CUDA, HIP from PI_KERNEL_GROUP_INFO_GLOBAL_WORK_SIZE
Fixes #8766
For instance (for X-dimension)
L0: maxGroupSizeX * maxGroupCountX
CUDA: CU_DEVICE_ATTRIBUTE_MAX_BLOCK_DIM_X * CU_DEVICE_ATTRIBUTE_MAX_GRID_DIM_X

[jinz2014]

I have a question. Is the max global work size independent of the global work size set in a host program for a kernel ?

[abagusetty]

/verify with intel/llvm-test-suite#1694

[bader]

@abagusetty, FYI. "verify with" command do not validate on CUDA/HIP platforms.
I recommend adding this test to sycl/test-e2e to validate on CUDA/HIP platforms.
+@aelovikov-intel.

[abagusetty]

Thanks, I stumbled upon that too and looked at the wording in Spec, which made me think it could be the max global limits.

The exact semantics of this descriptor are defined by each SYCL backend specification, but the intent is to return the kernel’s maximum global work size.

[jinz2014]

The global work sizes from the query will be the same for any kernels. Right ?

[abagusetty]

The global work sizes from the query will be the same for any kernels. Right ?

Yes, since the descriptor is a kernel_device_specific one: Any kernel from (custom device type or a built-in kernel) possibly returns the info of device specific global-work-sizes which in turn should be the same for all the kernels IMO.

[jchlanda]

It's a bit misleading, as it does not assert on the condition, maybe consider renaming it?

[abagusetty]

PI_ASSERT to PI_ERR_CHECK

[abagusetty]

Gentle ping @smaslov-intel @jchlanda

[jandres742]

+1 on L0 changes.

[steffenlarsen]

Sorry for the delay. I think these changes look good. I am a little curious what built-in kernels they would apply to, but I assume CUDA, HIP and L0 guarantee full possible work-sizes either way.

[abagusetty]

Sorry for the delay. I think these changes look good. I am a little curious what built-in kernels they would apply to, but I assume CUDA, HIP and L0 guarantee full possible work-sizes either way.

Thanks for the feed back on the built-ins, I too stumbled upon that a bit: Just convinced myself that they see the complete device limits.