DPC++ daily 2022-11-10

sycl-nightly/20221110

[SYCL][Docs] Move sycl_ext_oneapi_kernel_properties to experimental (…

DPC++ daily 2022-11-09

[SYCL][LIT] Fix narrowing type conversion for marray in test (#7298)

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

DPC++ daily 2022-11-08

[SYCL] Rename DoubleGRF to LargeGRF (#7284)

This change renames double GRF to large GRF both for users and
internally in the compiler. We're doing this because we got direct
feedback from customer facing engineers that we should use the large GRF
terminology, and it also makes the naming consistent with other compiler
work we are doing.

For the user,

```
set_kernel_properties(kernel_properties::use_double_grf);
```
still works, it will just throw a deprecated warning and will be removed
in a future release.

The new way is
```
set_kernel_properties(kernel_properties::use_large_grf);
```

There should be no ABI break because we still check the previous image
property name in the program manager, so applications built with an old
compiler work using the runtime from a new compiler. I confirmed this
with manual testing.

I will update the system test here to test the new flag as well:
https://github.com/intel/llvm-test-suite/blob/intel/SYCL/DeviceCodeSplit/double-grf.cpp

Signed-off-by: Sarnie, Nick <[email protected]>

DPC++ daily 2022-11-07

[SYCL][HIP] Add interop header and devcie specialization (#7145)

Fixes https://github.com/intel/llvm/issues/6635 for HIP.

Note that it does mean the HIP extension header must be included in user
applications in order to get the correct specialization, as is done in
the test:

```cpp
#include <sycl/ext/oneapi/backend/hip.hpp>
```

DPC++ daily 2022-11-06

[NFC][SYCL] Range reduction refactoring (#7180)

To make it look like delegating to NDRange version. I expect some future
change completely unify the paths, but this looks like a self-contained
atomic improvement.

DPC++ daily 2022-11-05

[SYCL] Enable aspect usage propagation pass and add diagnostics (#6982)

This commit adds a warning diagnostic for when there is a mismatch
between aspect usage propagated to a function and the function's
`device_has` attribute. Additionally, notes accompany the warning to
give the user a trace through to where the aspects originate. Since the
aspects are propagated in an LLVM pass, the warning is issued from
there, which means no source information is available by default. To
alleviate this, Clang CodeGen will now add srcloc metadata with an
encoded version of the source location to allow the backend to correctly
report the location for the various new diagnostics.

Additionally, this commit also adds the SYCLPropagateAspectsUsage pass
to the passes run for SYCL device code.

Signed-off-by: Larsen, Steffen <[email protected]>
Co-authored-by: Sabianin, Maksim <[email protected]>

DPC++ daily 2022-11-04

[SYCL][LIT] Remove sycl tests from check-all in Debug mode (#7262)

They are extremely slow for Debug compiler and sporadically timeout.
It would be still possible to run these through check-sycl target.

This is NFC for Github Actions CI since we don't use check-all and call
all targets separately.

DPC++ daily 2022-11-03

sycl-nightly/20221103

[NFC][SYCL] Replace reduCGFuncAtomic64 with reduCGFuncForNDRangeBothF…

DPC++ daily 2022-11-02

[SYCL] Improve mock PI plugin (#7198)

The patch adds more logic to default redefinitions of PI APIs:

    1. Allocate, refcount and deallocate various handles
    2. Handle sub-buffer creation

Also the patch adds support for adding PI functions to be
called in addition (before or after) to the original function.
This allows intercepting PI API calls for introspection while
still allowing original function take care of handles. Or add
some post processing of the returned values.

DPC++ daily 2022-11-01

[SYCL] Fix sycl-prof JSON output (#7128)

This patch fixes a problem related to the JSON output. Every element of
`"traceEvents"` has a comma appended, this is a problem in the case of
the last element, e.g.,
```
{
  "traceEvents": [ 
 ...
{"name": "piTearDown", ... , "ts": "16..81"}, <-- here is  
],
"displayTimeUnit":"ns"
}
```
This comma prevents the display of the JSON output in
`chrome://tracing/` or [speedscope](https://www.speedscope.app/) due to
its invalid format. This patch solves the problem by adding an empty
element before finalizing the JSON output.