[SYCL][UR][L0] Add UR_L0 environment variables

[jandres742]

Define UR_LO_DEBUG and UR_L0_SERIALIZE, and keep
their ZE_ counterparts for now.

And do some reorganization

[jandres742]

@smaslov-intel : please review.

[bmyates]

Seems we need a way to determine if this code is being executed as part of L0 PI or L0 UR adapter. One way to do it would be to add a compile define into cmake that is only set when compiling the adapater library, check it in this file and route calls accordingly.

[jandres742]

@bmyates: we could, but that would be short-termed since idea is that there's no PI+L0 anymore but PI+URL0, if I'm not mistaken. This is how the prints look like now with this change:

$ UR_L0_DEBUG=1 SYCL_PI_TRACE=2 <app>
...
---> piKernelRelease(
        <unknown> : 0x2a78f10
ZE ---> zeKernelDestroy(Kernel->ZeKernel)
UR ---> urProgramRelease(KernelProgram)
) --->  pi_result : PI_SUCCESS
...

[smaslov-intel]

Similarly, have UR_L0_SERIALIZE take precedence

[jandres742]

@smaslov-intel Rebased and introduced some of the suggestions.

[smaslov-intel]

@smaslov-intel Rebased and introduced some of the suggestions.

LGTM. Please avoid rebase if not strictly necessary. It allows for incremental reviews.

[jandres742]

@smaslov-intel : incremental changes on third commit: 69fd501

[jandres742]

Failing on CUDA with

********************
Failed Tests (1):
  SYCL :: GroupAlgorithm/reduce_sycl2020.cpp
failing also with CUDA. do we have a known issue for this also?

Seems like it is being fixed in #8860

[jandres742]

Rebased needed to include #8860

[bader]

@jandres742, could you resolve merge conflicts, please?

[jandres742]

@jandres742, could you resolve merge conflicts, please?

Thanks @bader . Rebased.

[jandres742]

@bader : this is failing with

********************
Failed Tests (1):
  SYCL :: AtomicRef/atomic_memory_order_acq_rel.cpp

it is flaky here #8847. is someone working on it? I will rerun to see if it works.

[jandres742]

only failure is still

********************
Failed Tests (1):
  SYCL :: AtomicRef/atomic_memory_order_acq_rel.cpp

Reported in #8847

@intel/llvm-gatekeepers : could we merge this, to avoid further delays and rebases?