Releases: uxlfoundation/oneMath
v0.6
Features in This Release
BLAS
- Add generic device, allowing portBLAS backend to run on any SYCL device.
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
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.
- Add support for cuSPARSE backend.
General updates
- Set
libonemkl.so
's rpath to$ORIGIN
allowing it to load the backend shared objects without settingLD_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.
Known Issues and Limitations
v0.5
Features in This Release
BLAS
- Improved support for ROCm backend
DFT
- Improve specification for compute function overloads, and implement in oneMKL interfaces. #503
- Add FWD/BWD STRIDES API: #514
Documentation update
- Added security.md file which includes guidance on reporting security incidences.
- Updated documentation and README with UXL Foundation details.
- Rewrite build documentation #510
Known Issues and Limitations
v0.4
Features in This Release
BLAS
- Improved support for rocBLAS and portBLAS backends.
- Enabled cuBLAS and rocBLAS backends with AdaptiveCpp compiler.
DFT
- Enabled portFFT backend for x86 CPU, Intel GPU, Nvidia GPU and AMD GPU for the DFT SYCL APIs
- Enabled the external workspace DFT SYCL API
Sparse BLAS
- Introduced Sparse BLAS interfaces with backend for x86 CPU and Intel GPU
Documentation update:
-
Replaced the usage of "Intel project for LLVM* technology" with the approved one "oneAPI DPC++ Compiler" in all documents
-
Readme file update:
- The tables of "Supported Configurations" were reorganized for better readability.
- Support of
Open DPC++
compiler removed for configurationsx86
+Intel(R) oneMKL
+BLAS
|LAPACK
- Updated the section of "Supported Operating Systems" and compiler version
Known Issues and Limitations
v0.3
Features in This Release
BLAS
-
Added NETLIB X86 CPU backend support in BLAS domain for DPC++, LLVM and hipSYCL compilers
-
Added Unified Shared Memory (USM) Interfaces support in BLAS domain
-
Extended support in BLAS domain with low-precision, mixed precision GEMMs and various batch implementations.
-
Extended support in BLAS domain with transposition routines.
-
Added cuBLAS Nvidia GPU backend support in BLAS domain for hipSYCL compiler
-
Added portBLAS X86 CPU, Intel GPU, Nvidia GPU and AMD GPU backend support in BLAS domain for DPC++ and LLVM compilers.
-
Added rocBLAS AMD GPU backend in BLAS domain for LLVM and hipSYCL compilers.
-
Extended MKLCPU and MKLGPU backend support in BLAS domain with pointers for scalar parameters.
-
Various updates in BLAS domain to align with SYCL 2020 spec and bug fixes
LAPACK
- Introduced LAPACK SYCL APIs support for x86 CPU and Intel GPU with the Intel® oneAPI Math Kernel Library, Nvidia GPU with the cuSOLVER library and AMD GPU with the rocSOLVER library.
DFT
- Introduced DFT SYCL APIs support for x86 CPU and Intel GPU with the Intel® oneAPI Math Kernel Library, Nvidia GPU with the cuFFT library and AMD GPU with the rocFFT library.
RNG
- Introduced RNG Device API support according to the oneAPI 1.3 specification
Service
-
Replaced FindMKL.cmake with MKLConfig.cmake from Intel oneMKL
-
Extended support for cmake fetchcontent module
Known Issues and Limitations
-
RNG Host API Philox4x32x10 engine may produce incorrect numbers in Skip_Ahead_Ex test in case of mklgpu backend and Intel oneMKL 2024.0 release
-
The LAPACK domain for x86 CPU with Intel(R) oneAPI Math Kernel Library backend has build failures with builds of the LLVM* compiler after 2023-09-27. Use builds from 2023-09-26 or earlier.