oneAPI DPC++ Compiler 2020-12

New features

SYCL Compiler

• Allow for multiple build options for opencl-aot [5e5703f]

SYCL Library

• Implement SYCL_INTEL_mem_channel_property
  extension [2f1f316]
• Add marray class as defined by SYCL 2020 provisional [5eebd1e]
• Implement dynamic batch size adjusting when using Level-Zero plugin
  [c70b047, cf0d053]
• Add online compilation API interface [70ac47d]

Documentation

• Proposal for new device descriptors extension
  was added [1ad813b]
• Added online compilation extension
  specification [e05a19c]

Improvements

SYCL Compiler

• (ESIMD only) Remove wrapping of buffer objects into images which caused
  problems like incorrect work of scatter/gather of 1- and 2-byte values
  [d2d20d6]
• Rename FPGA kernel attribute [[intel::stall_enable]] to
  [[intel::use_stall_enable_clusters]] [dab9deb]
• Add template parameter support for [[intel::max_global_work_dim]] and
  [[intel::no_global_work_offset]] attributes [bd8fcc7, a5fde5a]
• Remove partial-link path when dealing with fat static archives [f1aa7f4]
• Remove unused device library function definitions from linked program
  [e9423ff]
• Don't dump IR and dot files by default in the LowerWGScope [9d0e352]
• Support LLVM floating-point intrinsics in llvm-spirv and frontend
  [a5065ab]
• Add template parameter support for [[intel::no_global_work_offset()]]
  attribute [a5fde5a]
• Improve group size selection by adjusting parallel_for execution range
  global size [74a68b7]
• Add clang support for FPGA loop fusion function attributes [23926b0]
• Reword some compiler diagnostics [50b81c3]

SYCL Library

• Eliminate performance overhead on devices without host unified memory support
  [a4f0924]
• Optimize discard_write access mode for host accessor [6733c8b]
• Add support for composite specialization constants
  [c62860f, d4251e3, 3ec4594]
• Enhance PI tracing with printing output arguments [19f5ad6]
• Introduce pi_map_flags in lieu of cl_map_flags [f0e7606]
• Implement robust error handling in LevelZero plugin [65c719d]
• Add new device descriptors as SYCL extensions [51ac08c]
• Remove redundant dependencies for in-order queues [6327221]

Documentation

• Add information on AOT to GetStartedGuide [71942fb]
• Add notice on alignemnt checks in ABI policy [4326b95]
• Updated design of specialization contants on using of POD types
  [81963d1]
• Document linked allocation commands [929a764]
• Improve ESIMD documentation rendering [079597d]
• Improved device library documentation [f24e2a9]

Bug fixes

SYCL Compiler

• Do not customize optimizations for non-SPIR targets [cb069fe]
• Fix address space assertion with templates [8905a8c]
• Link libm-fp64 device library by default [ac93d6f]
• Add support for specialization constants' typenames declared in namespaces
  [f64f835]
• Fix loosing OpenMP device binary when program uses both OpenMP and SYCL
  offloading models [eb89f5e]

SYCL Library

• Add missing interoperability API to construct SYCL classes with Level-Zero
  handles [10b4e8a]
• Fix several builtins implementation for host device
  [8b82c67, 7867089]
• Fix possible hang upon application finish if streams were used [bd5893a]
• Fix failure when employing interoperability host task on queue constructed
  with reused context [9cff6c9]
• Fix "instantiation after specialization" warnings
  [56b9a1d, eadce94]
• Support copying of stream by value within a kernel without loss of output
  information [8d37cba]
• Fix handling of big and/or non-uniform work-groups in reduction kernels. The
  solution may change when reduction kernels precompilation/query approach is
  implemented [78e2599]
• Fix memory leak in event pool in Level Zero plugin [68fc780]
• Fixed issue with finalizing context of Level Zero plugin [6cfa921]
• Fix backend selection for SYCL_DEVICE_FILTER=* case [c54da15]
• Restore AccessorImplHost layout [a08eeb4]

Documentation

• Updated source checkout instruction for Windows in GetStartedGuide
  [9cde152]

API/ABI breakages

Known issues

• GlobalWorkOffset is not supported by Level Zero backend [6f9e9a7]
• The code with function pointer is hanging on Level Zero [d384295]
• User-defined functions with the same name and signature (exact match of
  arguments, return type doesn't matter) as of an OpenCL C built-in
  function, can lead to Undefined Behavior.
• A DPC++ system that has FPGAs installed does not support multi-process
  execution. Creating a context opens the device associated with the context
  and places a lock on it for that process. No other process may use that
  device. Some queries about the device through device.get_info<>() also
  open up the device and lock it to that process since the runtime needs
  to query the actual device to obtain that information.
• On Windows, DPC++ compiler enforces using dynamic C++ runtime for
  application linked with SYCL library by:
  • linking with msvcrt[d].dll when -fsycl switch is used;
  • emitting an error on attempts to compile a program with static C++ RT
    using -fsycl and /MT or /MTd.
    That protects you from complicated runtime errors caused by C++ objects
    crossing sycl[d].dll boundary and not always handled properly by different
    versions of C++ RT used on app and sycl[d].dll sides.
• The format of the object files produced by the compiler can change between
  versions. The workaround is to rebuild the application.
• Using cl::sycl::program API to refer to a kernel defined in another
  translation unit leads to undefined behavior
• Linkage errors with the following message:
  error LNK2005: "bool const std::_Is_integral<bool>" (??$_Is_integral@_N@std@@3_NB) already defined
  can happen when a SYCL application is built using MS Visual Studio 2019
  version below 16.3.0 and user specifies -std=c++14 or /std:c++14.
• Employing read sampler for image accessor may result in sporadic issues with
  Level Zero plugin/backend [2c50c03]
• Printing internal defines isn't supported on Windows [50628db]
• Group algorithms for MUL/AND/OR/XOR cannot be enabled for group scope due to
  SPIR-V limitations, and are not enabled for sub-group scope yet as the
  SPIR-V version isn't automatically raised from 1.1 to 1.3 [96da39e]
• We cannot run Dead Argument Elimination for ESIMD since the pointers to SPIR
  kernel functions are saved in !genx.kernels metadata [cf10351]