v0.6 release notes

[vmakovet]

Features in This Release

BLAS

LAPACK

RNG

DFT

Sparse BLAS

Documentation update

Known Issues and Limitations

[paveldyakov]

Please find release notes for RNG:

• Introduced Beta and Gamma distributions support in RNG Device API
• Introduced uint64_t type support for uniform distribution in RNG Device API
• Introduced (u)int8/int16 types support for Bernoulli distribution in RNG Device API

[Rbiessy]

My suggestion are:

General updates

• Set libonemkl.so's rpath to $ORIGIN allowing it to load the backend shared objects without setting LD_LIBRARY_PATH
• Use [ext_codeplay_enqueue_native_command].(https://github.com/intel/llvm/blob/bfc4015a3f858ef664e469a9a0dbda97c2d22a1f/sycl/doc/extensions/experimental/sycl_ext_codeplay_enqueue_native_command.asciidoc#differences-with-host_task) extension with all CUDA and ROCm backends, improving the synchronization mechanism.

BLAS

• Add generic device, allowing portBLAS backend to run on any SYCL device.

DFT

• Fix correctness issue with ROCm 6.0+.
• Allow portFFT backend to run on any SYCL device.

Sparse BLAS

• Update MKLCPU and MKLGPU backends to the new API following the latest oneAPI specification.

Note: it would be great if we can wait until the end of October to create this tag. We may be able to merge the PR https://github.com/oneapi-src/oneMKL/pull/527 to support cuSPARSE. We're still investigating some issues but aim to merge it by the end of October.