DPC++ daily 2022-09-08

[SYCL][Windows] Enable building lld by default on Windows (#6701)

The Windows-Clang CMake module uses `lld-link` by default, so
having lld built on windows is helpful for people wanting to build SYCL
applications using CMake.

Related to: https://github.com/intel/llvm/issues/6026

DPC++ daily 2022-09-07

[CI][NFC] Rename workflow job (#6572)

Just pure renaming of "resolve test matrix" to "generate test matrix"
Follow up on a comment made at https://github.com/intel/llvm/pull/6528

DPC++ daily 2022-09-06

[SYCL][CUDA] Fix get_native interop for device (#6649)

This patch fixes: https://github.com/intel/llvm/issues/6635

In https://github.com/intel/llvm/pull/6483, the implementation of `get_native` for device for the CUDA plugin was mistakenly moved to the experimental interface header, and so it was no longer available for the regular interface, causing build issues.

For the CUDA plugin there is currently two interfaces for the CUDA interop, the "legacy" one which is used by projects such as oneMKL and oneDNN, and the "experimental" one, defined in the `sycl/ext/oneapi/experimental/backend/cuda.hpp` header  which implements the interop as described in the CUDA backend specification proposed here: https://github.com/KhronosGroup/SYCL-Docs/pull/197

DPC++ daily 2022-09-03

[SYCL] Fix template argument propagation in AccessorSubscript (#6669)

Fixes an issue where partial subscripts would fail to compile due to template errors.

Signed-off-by: Larsen, Steffen <[email protected]>

DPC++ daily 2022-09-02

sycl-nightly/20220902

[SYCL][Doc] Fix doxygen generation: add SYCL_INLINE_VER_NAMESPACE to …

DPC++ daily 2022-09-01

[SYCL] Fix depends_on handling with pi commands (#5901)

Fixes two related issues:
1) if pi task is blocked by host task or host accessor it can not be enqueued and piEvent is not present in its event_impl. When we schedule new pi task with explicit (depends_on) dependency on the first one - its is absent in MDeps since we have no usual memory dependencies and present in MPreparedDepsEvents. MPreparedDepsEvents is used in enqueueImp for obtaining piEvents. Any events from MPreparedDepsEvents w/o pi events will be just skipped.
AddDep always call processDepEvent which distributes events to MPreparedDepsEvents (pi event expected) and MPreparedHostDepsEvents (no pi event) so replacement of MDeps in enqueueCommand should be valid.
2) if we have kernel w/o usual memory dependencies (MDeps & MUsers are empty) blocked kernel will be just "cleanup" and its execution will be skipped.

Signed-off-by: Tikhomirova, Kseniya <[email protected]>

DPC++ daily 2022-08-31

[Driver][SYCL] Add support for -fsycl-force-target (#6653)

Introduce -fsycl-force-target=arg support.  This is used along with -fsycl
to allow the user to override the target used to unbundle device objects
from fat objects and archives.

For example, object.o is built from -fsycl-targets=spir64.  The user wants
to build with -fsycl-targets=spir64_gen and object.o.  Use of
-fsycl-force-target=spir64 allows for this to be accomplished.

Additional notes:
  - Only valid when used with a single triple from -fsycl-targets
  - Applies to all unbundled archives and objects

DPC++ daily 2022-08-30

sycl-nightly/20220830

[SYCL][FPGA]Implementation of max_reinvocation_delay loop attribute (…

DPC++ daily 2022-08-29

[ESIMD] Fix saturation argument of DPAS (#6647)

* [ESIMD] Fix saturation argument of DPAS

The template argument for saturation was declared/used such a way
that any type could be passed to it, which would cause enforcement of
saturation when not intended.
In even worse scenarios the DPAS call with 3 simd arguments was
recognized as DPAS with 2 simd arguments + saturation argument:
dpas(src0,src1,src2) was treated as dpas(src1,src2,sat),
which caused totally incorrect behavior at runtime.

Also, this patch fixes the incorrect detection of ops_per_channel for
tfloat32 type on HOST.

Signed-off-by: Vyacheslav N Klochkov <[email protected]>

DPC++ daily 2022-08-27

[SYCL][SPIR-V] Change the LLVM type name of SPIR-V matrix types. (#6535)

This eliminates the need for the SPIR-V translator to query the pointer element
type of the members in the struct to figure out what matrix type it really is.