[SYCL][ESIMD] Report an error when slm_init is called more than once in the kernel

[fineg74]

No description provided.

[sarnex]

Can we send an email to the CFE/SYCL language team as a non-blocking follow-up to see if they have any ideas on improving this check by doing it somewhere else (Sema?). Relying on the metadata should work but it seems a bit indirect and possibly flaky, but I don't know how to do any better today.

[fineg74]

Will do, although it looks like a standard way to mark passing accessors to the kernel from FE

[v-klochkov]

I have couple of concerns in this patch: 1st - test is probably not valid (see other comment below),

2nd is related to a case with 1 kernel using slm_init and 2nd kernel using local_accessor.
Please add a test for this scenario.

[v-klochkov]

nit: the comment is not precise. It tests the case when a function is marked noinline, not not marked as always_inline.

[fineg74]

Fixed

[v-klochkov]

How about using such algorithm?

1. for each spir_function F or spir_kernel F available in the module have a map: map_func_to_esimd_kernel[F] = esimd_kernel
   Additionally:
   • If F has slm_init, then map_kernel_to_slm_num[map_func_to_esimd_kernel[F]]++;
   • If F is a kernel and uses local_accessor then map_kernel_to_slm_num[map_func_to_esimd_kernel[F]]++;. This ++ can be done only once per kernel (i.e. if 2 local accessors, then it is still 1 ++ op).
2. if F is not called by any esimd_kernel, then simply prohibit uses of slm_init for such F - report an error.
3. After the initial loop by F do walk through each of esimd_kernels in the map map_kernel_to_slm_num. The int value mapped there must not exceed 1.

There may be some extra data needed here to emit a meaningful/helpful error. It is more the idea above.

[fineg74]

How about using such algorithm?

1. for each spir_function F or spir_kernel F available in the module have a map: map_func_to_esimd_kernel[F] = esimd_kernel
   Additionally:

   • If F has slm_init, then map_kernel_to_slm_num[map_func_to_esimd_kernel[F]]++;
   • If F is a kernel and uses local_accessor then map_kernel_to_slm_num[map_func_to_esimd_kernel[F]]++;. This ++ can be done only once per kernel (i.e. if 2 local accessors, then it is still 1 ++ op).

2. if F is not called by any esimd_kernel, then simply prohibit uses of slm_init for such F - report an error.

3. After the initial loop by F do walk through each of esimd_kernels in the map map_kernel_to_slm_num. The int value mapped there must not exceed 1.

There may be some extra data needed here to emit a meaningful/helpful error. It is more the idea above.

1. An ESIMD function can be called from multiple kernels
2. Inside the kernel accessor is a pointer in some address space and currently the only way to identify it as an accessor is to examine the metadata when it is passed from host to the kernel. In other words, I do see a kernel using local accessors but I don't see a a particular function using a local accessor. (Technically it is possible to track whatever is passed to the kernel but I am not sure it worth the effort. I am also not sure if treating every pointer in the local address space as a local accessor (without using the kernel metadata) is a correct way to identify local accessors.

I believe the current implementation is using more or less the same algorithm but in slightly different way: It uses a set to see if a kernel called slm_init one way or another I probably need to tweak it for multiple kernels in a module

[v-klochkov]

Looks good. Thank you.

Just 1 minor comment - SPIRV_LOCAL_ACCESSOR_PREF is not needed after the recent logic simplification patch.