DPC++ daily 2022-03-08

[CODEOWNERS] Update with new names (#5743)

* Added Tianfei for Release Notes
* Added ESIMD reviewers for SYCLLowerIR/CMakeLists.txt

oneAPI DPC++ Compiler 2021-12

New features

SYCL Compiler

• Added support for -fgpu-inline-threshold which allows controlling inline
  threshold of the SYCL device code [5f7b607]
• Added experimental support for CUDA backend on Windows [8aa3513]
• Added support for experimental option -fsycl-max-parallel-link-jobs=<N>
  which can be used specify how many processes the compiler can use for
  linking the device code [c2221f0]

SYCL Library

• Added support for default context extension
  on Linux [315593d]
• Added experimental support for group sorting algorithm
  [932ae56]
• Added support for sub-group mask extension
  [78a3e77]
• Added sycl::ext::intel::experimental::esimd::simd_mask as a replaced for
  sycl::ext::intel::experimental::esimd::mask_type_t to represent Gen
  predicates [01351f1]
• Added stripped PDBs for SYCL libraries [6e5dd48]
• Added support ESIMD emulator backend [f4ad3c1]
• Added support for FPGA DSP control extension [790aa8b]
• Implemented discard_events extension
  [9542e28]
• Extended XPTI notifications with information about SYCL memory management
  [a068b15] [8f9d0d2]
• Added support for sycl::device::info::ext_intel_gpu_hw_threads_per_eu
  [2e798df]
• Added sycl::ext::oneapi::no_offset property for
  sycl::ext::oneapi::accessor_property_list [308e5ad]
• Implemented SYCL 2020 property traits [3b17da6]
• Implemented MAX_WORK_GROUP_QUERY extension
  [2fdf940]
• Added function wrapper for function-level loop fusion attribute
  ([[intel::loop_fuse(N)]] and [[intel::loop_fuse_independent(N)]]) for
  FPGA [e8ac5a0]
• Added querying number of registers for kernel with ext_codeplay_num_regs
  kernel info query [97d33b7]
• Added backend macros SYCL_BACKEND_OPENCL and SYCL_EXT_ONEAPI_* for CUDA,
  Level Zero, ESIMD emulator, HIP [2b0ebab]
• Added support for sycl::ext::intel::experimental::esimd_ballot function
  [0bbb091]
• Added initial support for Tensorcore matrix extension
  [711ba58]

Documentation

• Added device global extension specification
  [d3e70d4]
• Added property list extension specification
  [a7da8b4]
• Added extension specification for discard queue events
  [23ca24b]
• Added KernelProperties
  extension [64f5e70]

Improvements

SYCL Compiler

• Added diagnostics on attempt to pass an incorrect value to
  -fsycl-device-code-split [631fd69]
• Improved output of -fsycl-help[2404d02]
• Allowed ::printf builtin for CUDA backend only [0c55d3a]
• Implemented nextafter for sycl::half on CUDA backend [53c3268]
• Added atomics with scopes and memory orders for CUDA backend
  [2ebde5f] [00f43b3]
• Added support for missing mathematical builtins in CUDA backend
  [789ec8b] [f074774] [390e105]
• Added diagnostic for non-forward declarable kernel name types [653bae9]
• Added group_ballot intrinsic for CUDA backend [0680e5c]
• Added support for device side assert for CUDA backend [5a87b8c]
• Turned on -fsycl-dead-args-optimization by default [5983dfd]
• Improved compilation time by removing free function queries calls detection
  [e4791d1]
• Reduced memory consumption of device code linking [6266820]
• Improved UX of sycl::ext::oneapi::experimental::printf by allowing format
  string to reside in a non-constant address space [2d62e51]
• Improved barrier and sync instructions to use full mask when targeting NVPTX
  [5ce99b8]
• Added match for default SPIR device architecture with host architecture i.e.
  x86_64 matches spir64 and i686 matches spir [f4d01cd]
• Set default device code split mode to off for FPGA [bea72e6]
• Improved diagnostic for invalid SYCL kernel names
  [455dce8] [df1ff7a]
• Made Xsycl-target-frontend= to accept device tripple aliases [7fa0569]
• Improved diagnostic messages for -fsycl-libspirv-path [c54c605]
• Made implied default device to force emulation for FPGA builds [074944e]
• Added support for sycl::ext::oneapi::sub_group::get_local_id for HIP
  backend [7a9335d]
• Added a diagnostic of indirect implicit capture of this for kernel lambda
  [dce4c6a]

SYCL Library

• Updated joint matrix queries to report if unsigned int variants of mad
  matrix instruction are supported [dd7ebce]
• Reduced overhead of device code assert implementation [b94f23a] [58ac74e]
• Added a diagnostic on attempt to call sycl::get_kernel_id with an invalid
  kernel [9dd1ea3]
• Reduced overhead on kernel submission for CUDA backend [b79ae69]
• Reduced overhead on kernel submission in backend independent part of runtime
  [e292aa5]
• Aligned Level-Zero Interoperability API with SYCL 2020 specification
  [dd7f82c] [e662166]
• Made sycl::half default constructor constexpr [d32a444]
• Changed CUDA and HIP backends to report each device in a separate platform
  [8dddb11]
• Added initial support for SYCL2020 exceptions [15e0ab1]
• Added SYCL 2020 sycl::target::device enumeration value [f710886]
• Added a diagnostic on attempt to print std::byte using sycl::stream
  [dd5e094]
• Added possibility to specify ownership of ze_module_handle_t when creating
  a sycl::kernel_bundle from it [e3c9c92]
• Improve performance of sycl::nd_item::get_group_range() [0cd7b7e]
• Deprecated sycl::target::global_buffer
• Made device_num which can be passed to SYCL_DEVICE_FILTER unique
  [7aa5be0]
• Added a diagnostic on using mutually exclusive sycl::handler methods
  [6f620a4]
• Added support for std::byte to sycl::vec class [8fa04fe]
• Added sycl::make_kernel interoperability support for Level-Zero backend
  [98896fd]
• Optimized work with events in the Level Zero backend [973aee9]
• Added support for sycl::ext::oneapi::experimental::matrix::wi_slice and
  sycl::ext::oneapi::experimental::matrix::joint_matrix_fill
  [97127eb] [cbad428]
• Enabled code location information when NDEBUG is not defined in XPTI
  notifications [e9f2d64] [9ca7cea]
• Added a diagnostic on attempt to pass a command group function object to
  sycl::queue::single_task [2614d4d]
• Enlarged the maximum batch size to 64 for Level Zero backend to improve
  performance [596f693]
• Reduced kernel submission overhead for CUDA backend [35729a7]
• Improved translation of Level Zero error codes [6699a5d], [5d9a04b]
• Added support for an arbitrary number of elements to
  sycl::ext::intel::experimental::esimd::simd::copy_from/to methods
  [2bdc4c4]
• Added HIP support to sycl::ext::oneapi::filter_selector
  [7224cb2], [b7cee06]
• Added support for batching copy commands for Level Zero backend [4c3e699]
• Reduced sycl::queue::submit overhead by enabling post-enqueue execution
  graph cleanup [6fd6098]
• Added support for classes implicitly converted from sycl::item in
  sycl::handler::parallel_for parameter to align with the SYCL 2020
  specification [34b93bf]
• Removed direct initialization constructor from
  sycl::ext::intel::experimental::bfloat16 class [81154ec]
• Added sycl::vec and sycl::marray support to sycl::known_identity type
  trait [8fefb25]
• Added minimal support for the generic space address space to match
  sycl::atomic_ref class definition in specification [e99f298]
• Improved cache of command-lists in the context to be per-device for Level
  Zero backend [ca457d9]
• Extended group algorithms to support broadened types [3205368]
• Added support for alignement flags in
  sycl::ext::intel::experimental::esimd::simd::copy_from/copy_to operations
  [27f5c12]
• Made sycl::ext::oneapi::atomic_ref available in sycl namespace
  [2cdcbed]
• Renamed cuda and hip enum values and namespaces to ext_oneapu_cuda and
  ext_oneapi_hip to align with SYCL 2020 specification [97f916e]
• Improved performance of kernel submission process [535ad1e]
• Eliminated build of unwanted kernels when creating one with make_kernel
  [53ea8b9]
• Removed duplicate devices on submission to kernel_bundle API functions
  [c222497]
• Deprecated sycl::aspects::int64_base_atomics and
  sycl::aspects::int64_extended_atomics [554b79c]
• Made backend specific headers be included implicitly [bc8a00a]
• Removed program class and related API [e7cc7b0]
• Excluded current working directory from DLL search path when looking for
  runtime dependencies [0a65cb4]
• Enabled persistent device code cache for kernel bundles [810d67a]
• R...

DPC++ daily 2022-03-07

[BuildBot] Uplift GPU RT version for Linux to 22.09.22577 (#5742)

Signed-off-by: bb-sycl <[email protected]>

DPC++ daily 2022-03-05

sycl-nightly/20220305

[ESIMD] Disable non-critical messages from VC backend (disable-finali…

DPC++ daily 2022-03-04

[SYCL][ESIMD] Disallow use of accessor::operator[] in ESIMD code (#5706)

This operator is not supported in ESIMD context.

Signed-off-by: Sergey Dmitriev <[email protected]>

DPC++ daily 2022-03-03

sycl-nightly/20220303

[SYCL][FPGA] Expose value_type and min_capacity from SYCL pipes exten…

DPC++ daily 2022-03-02

[CI] Temporarily disable OCL FPGA Emulator (#5707)

FPGA emulator causes crashes on SYCL runtime initialization.

DPC++ daily 2022-03-01

[SYCL] Fix host device local accessor alignment (#5554)

Local kernel arguments must be aligned to the type size, simply using `std::vector<char>` doesn't always provide the correct alignment. So this patch adds extra padding to the vector and ensures that the pointer returned for the accessor is actually aligned to the type size.

This issue was exposed by: https://github.com/intel/llvm-test-suite/pull/608, which was a follow up to fixing local accessor alignment for the CUDA plugin.

DPC++ daily 2022-02-28

sycl-nightly/20220228

[SYCL][NFC] Fix warnings after compile-time properties implementation…

DPC++ daily 2022-02-27

sycl-nightly/20220227

[SPIR-V] Add SPV_INTEL_long_composites (#2848)