[SYCL] Implement sub-group mask extension

[vladimirlaz]

The specification is available under https://github.com/intel/llvm/blob/sycl/sycl/doc/extensions/SubGroupMask/SubGroupMask.asciidoc

Complementary test changes are available under intel/llvm-test-suite#441, intel/llvm-test-suite#462

[dm-vodopyanov]

SYCL RT changed LGTM

[vladimirlaz]

@intel/dpcpp-specification-reviewers, could you please have a look?

[gmlueck]

@intel/dpcpp-specification-reviewers, could you please have a look?

I added some comments just about the changes to the spec. It would be better to wait for a review from @Pennycook, though. He was involved more closely with this API, and he would be better to review the implementation. John is on vacation this week, returning on Monday.

[vladimirlaz]

@gmlueck, @rolandschulz, @Pennycook, could you please have a look on recent changes.
main changes:

• Change default value for max_bits to 32 (compiler flag will be added later)
• Remove marray<> from internal layout (leave for insert_bits and extract_bits). Use uint32_t to store bits.
• Make value for insert_bits(const &T in) and extract_bits(const &T out) deducible by putting in/out to function arguments.
• Rename to sub_group_mask.
• Update spec acordingly.

[Pennycook]

I think there are a few spec updates missing. GitHub won't let me comment on the diff for some reason, so doing so here:

The predefined macro section still says SYCL_EXT_ONEAPI_GROUP_MASK, whereas it should say SYCL_EXT_ONEAPI_SUB_GROUP_MASK after the renaming.

I thought we'd decided that we'd use marray<char, max_bits/CHAR_BIT> instead of marray<uint32_t, ...> for large masks, with a note that max_bits must be a multiple of CHAR_BIT. That would allow you to get rid of marray_size. If max_bits is small enough that you don't need an marray at all, the specification should allow for that too.

[vladimirlaz]

I think there are a few spec updates missing. GitHub won't let me comment on the diff for some reason, so doing so here:

The predefined macro section still says SYCL_EXT_ONEAPI_GROUP_MASK, whereas it should say SYCL_EXT_ONEAPI_SUB_GROUP_MASK after the renaming.

I thought we'd decided that we'd use marray<char, max_bits/CHAR_BIT> instead of marray<uint32_t, ...> for large masks, with a note that max_bits must be a multiple of CHAR_BIT. That would allow you to get rid of marray_size. If max_bits is small enough that you don't need an marray at all, the specification should allow for that too.

@Pennycook It looks like I forgot to upload the commits.

• SYCL_EXT_ONEAPI_GROUP_MASK has been renamed:
• I replaced marray<> with uint32_t as we have max_bits set to 32.
• I have removed marray_size from the sources and will remove from the doc (forgot it)

[vladimirlaz]

@Pennycook, @gmlueck, @dm-vodopyanov, @rolandschulz could you please review/approve the PR?

[Pennycook]

@Pennycook It looks like I forgot to upload the commits.

• SYCL_EXT_ONEAPI_GROUP_MASK has been renamed:

There's still one instance of SYCL_EXT_ONEAPI_GROUP_MASK (Line 54 of the extension documentation). Otherwise LGTM.

[vladimirlaz]

@Pennycook It looks like I forgot to upload the commits.

• SYCL_EXT_ONEAPI_GROUP_MASK has been renamed:

There's still one instance of SYCL_EXT_ONEAPI_GROUP_MASK (Line 54 of the extension documentation). Otherwise LGTM.

@Pennycook, I have fixed that. Could you please approve the PR?

[bader]

The specification is available under https://github.com/intel/llvm/blob/sycl/sycl/doc/extensions/SubGroupMask/SubGroupMask.asciidoc

@vladimirlaz, please, update the link to the specification.

[vladimirlaz]

The specification is available under https://github.com/intel/llvm/blob/sycl/sycl/doc/extensions/SubGroupMask/SubGroupMask.asciidoc

@vladimirlaz, please, update the link to the specification.

the spec is moved in the scope of the PR. The comment contains the new location.