Intel® oneAPI Base Toolkit Release Notes

Features at toolkit level

• macOS* support for CPU target with 8 components, see details in Intel oneAPI Base Toolkit System Requirements

• Intel® C++ Compiler is now included with icx clang based driver

• the default installation path is changed:
  • Linux or macOS: /opt/intel/oneapi
  • Windows: C:\Program Files (x86)\Intel\oneAPI

Intel® oneAPI DPC++ Compiler & Intel® C++ Compiler

• Intel® oneAPI DPC++ Compiler
  • Performance improvement of sycl_benchmarks on CPU
  • Manual lsu control
  • Support equivalent of CL_CHANNEL_*_INTELFPGA and forward to clCreateBuffer call
• Intel® C++ Compiler is now included with following features:
  • icx clang based driver
  • vectorization unrolling
  • OpenMP* SIMD ORDERED
  • Initial OpenMP* 5.0 support

Intel® oneAPI DPC++ Library

• RNG package integration
• parallel_transform_scan refactoring
• Public interface refactoring
• Parallel_sort optimization for non arithmetic types and operators that are not '<', '>'
• Range based API support.
• Performance optimization

Intel® DPC++ Compatibility Tool

• Nvidia* Thrust* migration coverage improved.
• More migration coverage for CUDA* Libraries (cuBLAS, cuSPARSE, cuSolver, and cuRand)
• Updates to DPCT header files (dpct::)

Intel® oneAPI Math Kernel Library

• Added DPC++ support for Host, CPU & GPU for select functions.
• Added DPC++ support for Host & CPU for select functions.
• Added MKL Graph support for GraphBLAS
• Added support for Windows DPC++ dynamic linking

Intel® oneAPI Threading Building Blocks

• Rework (revamp) TBB to align with modern C++
• Added NUMA API support
• Remove deprecated functionalities
• Improvements in TBB malloc and scheduler

GDB*

• Support C++ OpenMP* offload debugging on Linux*.
• Support non-stop mode to inspect a stopped GPU thread while the other CPU & GPU threads are running. The application debugger does not stop the other threads in non-stop mode.
• Converging the CPU debug capabilities including
  • Intel(R) Processor Trace Write (ptwrite),
  • Visual Function Call History (func_call),
  • Transactional Synchronization Extensions (TSX),
  • BFLOAT16, etc.

Intel® Integrated Performance Primitives

• Added IPP Cryptography domain
• Extended optimization for Intel® IPP Cryptography cipher AES, RSA support on 10th Generation Intel® Core™ processor family.
• Added new universal CRC function to compute CRC8, CRC16, CRC24, CRC32 checksums
• Reinstated the ippiComplement function optimized for Intel® AVX-512, AVX2, SSE4.2

Intel® oneAPI Collective Communications Library

• GPU scale-out
• Sparse buffers
• Horovod on Spark

Intel® oneAPI Data Analytics Library

• Added new APIs for PCA, K-means, KNN, Random Forest, SVM algorithms
• Added full oneTBB intergrated support
• Performance optimizations

Intel® oneAPI Deep Neural Networks Library

• Extended GEN12LP optimizations
• Initial int8 SPR optimizations
• Extended eltwise primitive with support for 'round' operation.
• GRU (lbr) for GPU

Intel® oneAPI Video Processing Library

• New oneVPL library and reference implementation with backwards compatibility to Intel® Media SDK
• CPU plugin implementation with H.265 & AV1 SW Decode/Encode
• oneVPL dispatcher with C API
• First release to the Open Source repository

Intel® Distribution for Python*

• Initial support for Intel GPU execution of Python/Numba code on Linux OS.
• Initial release for PyDPPL a Python wrapper for SYCL and OpenCL.
• Releasing mkl_sparse package for MKL-powered sparse matrices and mkl_umath package for Intel® technologies-powered NumPy universal function.

Intel® Advisor

• Added memory-level Roofline analysis, that helps to pinpoint exact memory hierarchy bottlenecks (L1, L2, L3 or DRAM)
• Technical Preview: New and improved Intel(R) Advisor user interface workflows and toolbars, incorporating Roofline analysis for GPUs and Offload Advisor
• Technical Preview: Offload Advisor introduces new data transfer metrics, hinting on potential bottlenecks.

Intel® VTune™ Profiler

• Refining analysis for GPU accelerators with the addition of OpenMP* offload pragna-aware metrics to HPC performance analysis.
• Added a Performance Snapshot as a first profiling step to suggest the detailed analyses (memory, threading, etc.) that offer the most optimization opportunity.
• Level-0 support on Windows

Intel® FPGA Add-On for oneAPI Base Toolkit (Optional)

• New Custom platform add-on containing a newer quartus version

Features at toolkit level

• Level-0 runtime is supported and enabled on Windows for some components
• "module" support on Linux*

Intel® oneAPI DPC++ Compiler

• Complex data types support
• Compiler switch to Level-0 by default on Windows
• C++17 is enabled by default

Intel® oneAPI DPC++ Library

• Implemented "device_vector" and "host_vector"
• Implemented c99 complex APIs for accelerators on Windows

Intel® DPC++ Compatibility Tool

• Migration improvement on Math and Texture/Surface API calls
• Thrust* migration improvement
• Updates to dpct::headers

Intel® oneAPI Math Kernel Library

• Level-0 is supported by default on Windows*

For GPU:

• RNGs – Multiple functions added.
• Vector Math - USM support for device-only pointers.

For Host, CPU & GPU:

• BLAS - strided batch AXPY
• FFT – 2D & 3D R2C OpenCL kernels
• LAPACK – USM interface for STEQR.

For Host, CPU & GPU:

• LAPACK – Multiple pointer-based USM interfaces to functions
• Summary Statistics – Multiple functions added.

GDB*

• Support debugging kernels using OpenMP offload directives
• Support for SIMD lane syntax for the "thread apply" commands

Intel® oneAPI Threading Building Blocks

• No change

Intel® Integrated Performance Primitives

• Bug fixes

Intel® oneAPI Collective Communications Library

• Enabled Deep learning frameworks on GEN9

Intel® oneAPI Data Analytics Library

• Multiple new GPU Machine Learning algorithms
• Several CPU performance optimizations
• Level-0 is supported as default on Windows

Intel® oneAPI Deep Neural Networks Library

• Support for Arm* 64-bit architecture (AArch64) and other non-x86 processors
• Provided APIs for primitive cache management
• Level-0 is the default on Windows

Intel® oneAPI Video Processing Library

• Expanded codecs (AV1, VP9 decode+encode)
• API enhancements on Composition, Async, Sharpen, procamp, and JSON parameters

Intel® Advisor

Offload Advisor:

• support GPU performance projection on Windows host
• capable of detecting data transfer of stack-allocated objects

Intel® Distribution for Python*

• Release scikit-ipp 1.0.0 for Image warping , image filtering and morphological operations.
• Implemented performance optimizations for K-Means algorithm in scikit-learn
• Initial support of scikit-learn with the device context (GPU)

Intel® VTune™ Profiler

• Intel® VTune™ Profiler now supports the latest Intel GPUs including Gen9, Gen11, Gen12
• Added an improved GPU Memory Hierarchy diagram annotated with GPU hardware performance metrics.
• I/O analysis is improved with a better summary and data presentation with additional DDIO metrics for Sky Lake and Cascade Lake servers.
• Level-0 support on Linux*

Intel® FPGA Add-On for oneAPI Base Toolkit (Optional)

• no change

Features at toolkit level

• Added command prompt menu item for each Microsoft Visual Studio* edition installed
• Level-Zero RT support for GPU on Linux*

Intel® oneAPI DPC++ Compiler

• Compiler switch to Level-zero by default on Linux*
• Limited interoperability between Level-zero and OpenCL*
• Specialization constants

Intel® oneAPI DPC++ Library

• Support parallel_sort (radix sort) optimization
• support std::complex on Windows*

Intel® DPC++ Compatibility Tool

• USM-enabled migration of cuRAND API calls (Note not all API calls can be migrated by the tool)
• More memory management API migration coverage, improved stream management API migration
• Updates to dpct:: headers

Intel® oneAPI Math Kernel Library

• Unified Shared Memory (USM) support for limited set of LAPACK functions
• DPC++ functionality (only USM interfaces) with CPU/GPU support for HPCG benchmark
• DPC++ functionality with GPU support for MCG31 and MCG59 engines

GDB*

• Support Intel integrated graphics GEN11

Intel® Integrated Performance Primitives

• Support BZIP2 v1.0.8
• Optimize Resize 8u for Ice Lake
• Develop ippsFIRSparse_32fc

Intel® oneAPI Collective Communications Library

• oneDAL on Spark* enabling
• BigDL enabling

Intel® oneAPI Data Analytics Library

• DBSCAN algorithm
• SVM algorithm

Intel® oneAPI Deep Neural Networks Library

• LSTM with projection for CPU
• Eigen threadpool support for CPU
• Level-zero RT support for GPU

Intel® oneAPI Video Processing Library

• Expanded codecs (10bit SW encode, JPEG/MJPEG)
• API enhancements

Intel® Advisor

• Offload Advisor: support GPU performance projection on Windows host

Intel® Distribution for Python*

• Decision Tree Classifier & prediction in scikit-learn
• Adaboost ensembles for Scikit-learn

Intel® VTune™ Profiler

• System Overview HW tracing Improvements: represent module entry points, user/kernel metrics, interrupts
• Improvement on proper handling of too short collections in case of PMU-based analysis

Intel® FPGA Add-On for oneAPI Base Toolkit (Optional)

• Support 3 FPGA boards (including PAC A10, PAC S10 and custom platform) with 3 add-on installers.

Features at toolkit level

• Automatic detection of supported GPU driver and print out a warning message if the minimal required version of GPU driver is not installed.
• If beta04 or older beta03 is installed on the system, please first uninstall beta04 or beta03 by following the Intel® oneAPI Toolkit Installation Guide and install the beta05 after. The side-by-side installation support is not fully implemented in this beta05 release.
• Two new plugins for Microsoft* Visual Studio Code* (VS Code) (one Sample Browser plugin for Intel® oneAPI Toolkits and one Launcher plugin for Intel® oneAPI Analyzers) are available on VS Code Marketplace at https://marketplace.visualstudio.com/publishers/intel-corporation

Intel® oneAPI DPC++ Compiler

• Initial multi-tile GPU management
• Device code split on modules
• FPGA diagnostic improvements
• Basic DPC++ & OpenMP* composability

Intel® oneAPI DPC++ Library

• Improved support for Unified Shared Memory (USM) pointers
• Extension APIs to support USM version of device_ptr

Intel® DPC++ Compatibility Tool

• Updated to dpct::headers
• Migration of cuRAND APIs to oneMKL RNG improved
• USM-enabled migration of BLAS API

Intel® oneAPI Math Kernel Library

Introduced DPC++ support for the following:

• For Host, CPU & GPU:
  • Triangular system solver (USM interfaces)
  • Input data to all Sparse BLAS functionality (USM interfaces)
  • Sparse GEMM function (buffer-based and USM interfaces)
• For Host & CPU:
  • Select LAPACK functions (USM interfaces)
  • Non-deterministic random number generator
• For GPU:
  • Orthogonal & unitary matrix multiplication (buffer-based interfaces)

Added the following OpenMP GPU offload features:

• Asynchronous offload capability for BLAS functions
• Triangular system solver (TRTRS)

Intel® oneAPI Collective Communications Library

• Added bfloat16 datatype support (except ccl_sparse_allreduce)
• Added OFI/SHM provider support
• Added Alltoallv collective

Intel® oneAPI Data Analytics Library

New CPU functionality:

• Elastic Net algorithm with L1 and L2 regularization in batch computation mode.
• Probabilistic classification for Decision Forest Classification algorithm with a choice voting method to calculate probabilities.

Intel® oneAPI Deep Neural Networks Library

• No new features

Intel® oneAPI Video Processing Library

• New features added:
  • Encode() to CPU extension
  • SVT-HEVC Encoding
  • 10-bit Decode support (CPU, GEN)
• Enhanced VPL Memory

Intel® Advisor

• GTpin 2.6 integration
• First Bottleneck visualization for Integrated Roofline chart
• Integrated Roofline single kernel view and guidance
• Bug fixes

Intel® Distribution for Python*

• daal4py accelerates DBSCAN algorithm in Scikit-Learn.
• Added support for Elastic net algorithm and Classification probabilities for Decision Forest algorithm in daal4py.

Intel® VTune™ Profiler

• Easier workflow for FPGA performance analysis
• Additional GPU performance metrics

Intel® FPGA Add-On for oneAPI Base Toolkit (Optional)

• no change

Features at toolkit level

• In this release Intel® oneAPI Base Toolkit supports co-existence with Intel® Parallel Studio XE or Intel® System Studio on Windows* from command line or in Visual Studio 2017* or Visual Studio 2019*.
• GPU driver for Ubuntu OS is no longer in the Intel® oneAPI Base Toolkit package; please read the Installation Guide for Intel® oneAPI Toolkits for installation instructions of the GPU driver.

Intel® oneAPI DPC++ Compiler

• Visual Studio integration supports Intel® Performance Libraries (oneMKL, oneTBB, IPP and oneDAL).
• "printf" support in device code
• Introduced the infrastructure on supporting standard C/C++ libraries in device code.
• queue::submit now throws synchronous exceptions
• cl::sycl::pipe class moved to cl::sycl::intel namespace
• Added single_task and parallel_for methods to cl::sycl::ordered_queue::single_task.
• Bug fixes, diagnostic improvements and other enhancements. See more in Intel® oneAPI DPC++ Compiler Release Notes.

Intel® oneAPI DPC++ Library

• Added 64-bit atomics support.
• Added <complex>, most functions in <cmath> GNU libstdc++), and <ratio>.
• Bug fixes and enhancements. See more in Intel® oneAPI DPC++ Library Release Notes.

Intel® DPC++ Compatibility Tool

• More texture API migration coverage
• More memory management API migration coverage
• Added the migration of 14 cuRAND APIs to oneMKL RNG
• Many improvements on readability and maintainability in generated code
• Bug fixes and improvements. See more in Intel® DPC++ Compatibility Tool Release Notes.

Intel® oneAPI Math Kernel Library

• BLAS: pointer-based (USM) interface support; C/C++ OpenMP* GPU offload support.
• LAPACK: pointer-based (USM) interface support for PORTF and POTRS; C/C++ OpenMP GPU offload support for xGETRF; GPU device support for POTRI.
• Sparse: Optimized solver of triangular system (TRSV) for matrices in CSR for GEN9.
• FFT: added support for C/C++ OpenMP offload to 1D C2C FFTs (single/double precision).

Intel® oneAPI Collective Communications Library

• Support for RHEL 7.x (CPU only) and CentOS 7.x (CPU only)
• Added 2D and pipelined ring allreduce algorithms (CCL_ALLREDUCE=2d, CCL_ALLREDUCE=ring).
• Bug fixes, stability and performance improvements.

Intel® oneAPI Data Analytics Library

• Initial support for heterogeneous input data formats
• Initial support for Unified Shared Memory (USM) for Homogeneous data
• New GPU single-node algorithms, i.e. Gradient Boosted Trees, Linear Regression, K-nearest Neighbors etc.
• Bug fixes, stability improvement.

Intel® oneAPI Deep Neural Networks Library

• New supported primitives: Matmul, LogSoftmax and Resampling;
• New functionality support: bfloat16 to RNN on GPU; asymmetric quantization support for MLPs
• Integration with linux-perf profiling tool
• Performance imporvements on CPU and GPU
• Bug fixes

Intel® oneAPI Video Processing Library

• Basic Encode (AVC, HEVC) for GEN9
• Transcode (AVC, HEVC)
• Enhanced VPL Memory; Improved extension design and Decode (AVC, HEVC)

Intel® Advisor

• GPU roofline collection supports Windows in this release.
• Migrated to Python 3.*
• Enhancement and bug fixes.

Intel® Distribution for Python*

• Optimized sklearn.cluster.DBSCAN using DAAL for automatic and brute force methods.

Intel® VTune™ Profiler

• Simplified system configuration requirements for GPU analysis. GPU utilization analysis is now available without a prerequisite of rebuilding the Linux kernel. See more in VTune Profile Release Notes below.
• The System Overview analysis type contains CPU and GPU Concurrency analysis.
• Enhancement and bug fixes.

Intel® FPGA Add-On for oneAPI Base Toolkit (Optional)

• Improved the installation script to work with Ubuntu OS and to remove duplicated library files.
• Resolved a license check issue due to expired license that caused the FPGA Add-On to stop working in year 2020.

• In this release the following features are provided:
  • Windows* 10 x64 is supported. FPGA support on Windows is limited to emulator only.
  • Integration of Microsoft* Visual Studio* 2017 and 2019
• Intel® oneAPI DPC++ Compiler supports SYCL specification 1.2.1 revision 5 plus extensions, i.e.:
  • Unified Shared Memory - Explicit and Restricted capabilities are supported
  • Sub-groups for NDRange Parallelism
• Intel® DPC++ Compatibility Tool supports:
  • Partial migration of kernel definitions/calls, memory management (including unified memory), device management, data types, error handling, math, events, streams, and more.
• GDB* supports:
  • Breakpoints inside a kernel
  • Source-level stepping
• Intel® oneAPI Math Kernel Library (oneMKL) supports:
  • BLAS, LAPACK, FFT, Sparse, Vector math, and Random number generator
• Intel® oneAPI Data Analytics Library supports:
  • K-Means Clustering
  • Principal Components Analysis (PCA)
• Intel® oneAPI Deep Neural Network Library provides:
  • Support for DPC++ compiler
  • SYCL API extensions and interoperability with SYCL code