0.21.0 Clean House

Added

• Support addition and multiplication-based prange reduction loops (#999)
• Proper boxing, unboxing of dpctl.SyclQueue objects inside dpjit decorated functions (#963, #1064)
• Support for queue keyword arguments inside dpnp array constructors in dpjit (#1032)
• Overloads for dpnp array constructors: dpnp.full (#991), dpnp.full_like (#997)
• Support for complex64 and complex128 types as kernel arguments and in parfors (#1033, #1035)
• New config to run the ConstantSizeStaticLocalMemoryPass optionally (#999)
• Support for Numba 0.57 (#1030, #1003, #1002)
• Support for Python 3.11 (#1054)
• Support for SPIRV 1.4 (#1056, #1060)

Changed

• Parfor lowering happens using the kernel pipeline (#996)
• Minimum required Numba version is now 0.57 (#1030)
• Numba monkey patches are moved to numba_dpex.numba_patches (#1030)
• Redesigned unit test suite (#1018, #1017, #1015, #1036, #1037, #1072)

Fixed

• Fix stride computation when unboxing a dpnp array (#1023)
• Using cached queue instead of creating new one on type inference (#946)
• Fixed bug in reduction mul operation for dpjit (#1048)
• Offload of parfor nodes to OpenCL UHD GPU devices (#1074)

Removed

• Support for offloading NumPy-based parfor nodes to SYCL devices (#1041)
• Removed rename_numpy_functions_pass (#1041)
• Dpnp overloads using stubs (#1041, #1025)
• Support for like keyword argument in dpnp array constructor overloads (#1043)
• Support for NumPy arrays as kernel arguments (#1049)
• Kernel argument access specifiers (#1049)
• Support for dpctl.device_context to launch kernels and njit offloading (#1041)