[ESIMD] Replace mask_type_t with simd_mask to represent Gen predicates.

[kbobrovs]

Addresses the following review comment from simd.hpp:
// TODO @rolandschulz, @mattkretz
// Introduce simd_mask type and let user use this type instead of specific
// type representation (simd<uint16_t, N>) to make it more portable
// TODO @iburyl should be mask_type_t, which might become more abstracted in
// the future revisions.
//

Gen predicates (masks) used to be represented with simd<unsigned short, N> types
which made them indistinguishable from simd objects forcing the same same sets
of available APIs, even though it did not always make sense. This patch
introduces a new class simd_mask, which provides its own set of APIs, some of
which are unique, other - repeat those from the simd class.

A new class 'simd_obj_impl' is added, which encompasses functionality common
between simd and simd_mask objects. Particularly, view creation APIs (select,
bit_cast_view), object construction APIs - replicate, memory I/O (copy from/to)
and others. simd and simd_mask both extend the simd_obj_impl class. They use the
'curiously recurring template' design pattern, when simd_obj_impl is templated
by its derivative class (simd or simd_mask), to allow simd_obj_impl APIs return
subclass objects rather than simd_obj_impl instances, those avoiding API
duplication in subclasses. For example, 'select' functions defined in
simd_obj_impl, will return a view of the derived class (subclass) rather than
a view of simd_obj_impl: simd_view<Derived,...>, where Derived is the derived
class.

APIs unique to simd_mask:

• unary logical negation '!' operator
• implicit conversion to bool for single-element masks
• logical binary operators '||' and '&&'

APIs unique to simd:

• binary arithmetic operators ('+', '-', '*', '/') and corresponding compound
  assignments
• unary arithmetic operators ('+', '-', '++', '--')
• relational operators ('<', '<=', '>', '>=')

The same division of the APIs to unique sets applies to simd_view<simd<...>,...>
and simd_view<simd_mask<...>,...> simd_view class specializations.

Complementary test change: intel/llvm-test-suite#388

Signed-off-by: kbobrovs [email protected]

[kbobrovs]

Rebased and resolved conflicts.

[kbobrovs]

Jenkins/Precommit failure is because required test fixes intel/llvm-test-suite#388 are not committed yet.

[DenisBakhvalov]

I spent some time reviewing it, but before going into more details, let me ask a few high-level questions:

• do we need simd_view from simd_mask? I noticed a few places where you distinguish whether a simd_view originates from simd or from simd_mask
• The patch is quite big. I think it may be split in the following parts:
  1. introduce simd_obj_impl class
  2. helper patch with fixes for simd_view class, changes in types.hpp, etc.
  3. operators
  4. introduce simd_mask class

[kbobrovs]

I spent some time reviewing it, but before going into more details, let me ask a few high-level questions:

• do we need simd_view from simd_mask? I noticed a few places where you distinguish whether a simd_view originates from simd or from simd_mask

Yes, we do. The reason is that simd_view<simd_mask,...> and simd_view<simd,...> support different set of APIs.

• The patch is quite big. I think it may be split in the following parts:

  1. introduce simd_obj_impl class
  2. helper patch with fixes for simd_view class, changes in types.hpp, etc.
  3. operators
  4. introduce simd_mask class

Unfortunately, it is next to impossible to split this into separate patches where each can pass building and testing individually.

[bader]

@kbobrovs, please, resolve merge conflicts.

[kychendev]

@kbobrovs, please double check there is no code quality regression with this change set. Thanks.

[kbobrovs]

@kbobrovs, please double check there is no code quality regression with this change set. Thanks.

@kychendev, in the absence of the perf regression suite (TBD), I manually checked linear, kmeans, bitonic sort v2, mandelbrot and dgetrf - no regressions.

[kbobrovs]

Precommit failed due to old version of llvm-test-suite, fixed by intel/llvm-test-suite#388.

@kychendev, @DenisBakhvalov - please approve

[kbobrovs]

@kychendev, @DenisBakhvalov - I manually checked CG (unfortunately, objcopy __CLANG_OFFLOAD_BUNDLE__sycl-spir64 somehow produces invalid .spv now - will file a separate JIRA), and caught one problem indeed - now simd<...>{val} invokes longer initializer list-based constructor as opposed to (val, stride) version. Fixed. Please take another look.

[kychendev]

@kychendev, @DenisBakhvalov - I manually checked CG (unfortunately, objcopy __CLANG_OFFLOAD_BUNDLE__sycl-spir64 somehow produces invalid .spv now - will file a separate JIRA), and caught one problem indeed - now simd<...>{val} invokes longer initializer list-based constructor as opposed to (val, stride) version. Fixed. Please take another look.

Thanks for double checking code quality and catching the issue. Added few comments.

[kbobrovs]

@kychendev, @DenisBakhvalov - review comments addressed.

[kbobrovs]

Jenkins failures are because the test suite is not updated yet. intel/llvm-test-suite#388 is the fix.

[2021-09-21T19:33:41.227Z] Failed Tests (14):
[2021-09-21T19:33:41.227Z]   SYCL :: ESIMD/BitonicSortK.cpp
[2021-09-21T19:33:41.227Z]   SYCL :: ESIMD/BitonicSortKv2.cpp
[2021-09-21T19:33:41.227Z]   SYCL :: ESIMD/PrefixSum.cpp
[2021-09-21T19:33:41.227Z]   SYCL :: ESIMD/Prefix_Local_sum1.cpp
[2021-09-21T19:33:41.227Z]   SYCL :: ESIMD/Prefix_Local_sum2.cpp
[2021-09-21T19:33:41.227Z]   SYCL :: ESIMD/Prefix_Local_sum3.cpp
[2021-09-21T19:33:41.227Z]   SYCL :: ESIMD/Stencil.cpp
[2021-09-21T19:33:41.227Z]   SYCL :: ESIMD/histogram_raw_send.cpp
[2021-09-21T19:33:41.227Z]   SYCL :: ESIMD/kmeans/kmeans.cpp
[2021-09-21T19:33:41.227Z]   SYCL :: ESIMD/mandelbrot/mandelbrot.cpp
[2021-09-21T19:33:41.227Z]   SYCL :: ESIMD/mandelbrot/mandelbrot_spec.cpp
[2021-09-21T19:33:41.227Z]   SYCL :: ESIMD/reduction.cpp
[2021-09-21T19:33:41.227Z]   SYCL :: ESIMD/stencil2.cpp
[2021-09-21T19:33:41.227Z]   SYCL :: ESIMD/usm_gather_scatter_rgba.cpp