1. OneAPI industry plan
1.1 Programming Challenges for Multiple Architectures
Increasingly specialized workloads require a variety of data-centric hardware. Today, each architecture requires a separate programming model and toolchain. The complexity of software development limits freedom of architecture choice.
1.2 OneAPI A programming model for multiple architectures and vendors
Freedom of choice
Choose the best acceleration technology that software cannot decide for you
Realize all hardware value
Performance across CPUs, GPUs, FPGAs, and other accelerators
Develop and deploy software with confidence
Open industry standards provide a safe, clear path to the future Path
▪ Compatible with existing languages and programming models, including C++, Python, SYCL, OpenMP, Fortran and MPI
1.3 Powerful OneAPI library
Realizes all hardware value
Designed to accelerate domain-specific critical function
freedom of choice
Pre-optimized for each target platform for maximum performance
2. Intel oneAPI products
Based on Intel's rich CPU tool heritage and extended to the XPU architecture, a complete set of high-level compilers, libraries, and porting, analysis, and debugger tools Utilizes state-of-the-art
hardware capabilities to accelerate computing
Compatible with existing programming models and code bases (C++, Fortran, Python, OpenMP, etc.) interoperability, developers can rest assured
that existing applications will work seamlessly with oneAPI
▪ Easy transition to new systems and accelerators - using a single code base gives developers more time to innovate
2.1 Intel® DPC++ Compatibility Tool
Assist developers in migrating code written in CUDA to DPC++ in one go, generating human-readable code as much as possible
About 90-95% of the code is usually automatically migrated
Provides inline comments to help developers complete application porting
2.2 Analysis and debugging tools
2.3 Intel® oneAPI Toolkit
A suite of proven developer tools that scales from CPUs to more accelerators
3. Details about Intel® oneAPI Toolkit
3.1 Intel® oneAPI Basic Toolkit
A core set of tools and libraries for developing high-performance applications on Intel® CPUs, GPUs, and
FPGAs.
For whom?
▪ Various developers in different industries
▪ Add-on tool suite users as it is the foundation of all tool suites
Key features/benefits
▪ Data-parallel C++ compiler, library and analysis tools
▪ DPC++ compatibility tools to help migrate existing applications written in CUDA With code
▪ Python distribution includes accelerated scikit-learn, NumPy and SciPy libraries
▪ Optimized performance libraries with support for threading, math, data analysis, deep learning and
video/image/signal
3.2Intel ® oneAPI DPC++/C++ Compiler
The compiler delivers exceptional parallel programming efficiency and performance across CPUs and accelerators
▪ Supports code reuse for different target hardware, as well as custom tuning for specific accelerators
▪ Open, cross-industry alternative to proprietary languages
DPC++ is based on ISO C++ and Khronos SYCL
▪ Uses common and familiar C and C++ constructs to achieve the efficiency benefits of C++
▪ Integrates with Khronos Group's SYCL to support data parallelism and heterogeneous
programming years of experience
3.3 Intel® oneAPI converged communication library optimizes communication mode
- Provides an optimized communication mode to achieve high performance on Intel CPUs and GPUs to distribute model training to multiple nodes
- Transparently supports multiple interconnects such as Intel® Omni-Path Architecture, InfiniBand, and Ethernet
- Built on low-level communication middleware (MPI and libfabrics)
- Support for efficient implementation of ensembles for deep learning training — all-gather, all-reduce, and reduce-scatter
3.4 Intel® VTune™ Analyzer
DPC++ analysis⎯for CPU, GPU and FPGA tuning
Profile Data Parallel C++ (DPC++)
View the most time-consuming lines of DPC++ code Tuning
for Intel CPUs, GPUs, and FPGAs Optimizing offloading
for any supported hardware accelerator Tweaking OpenMP offloading performance Various performance profiles CPU, GPU, FPGA, threads, Memory, Cache, Storage… Support for common languages DPC++, C, C++, Fortran, Python, Go, Java or a combination of languages
3.5 Intel® Advisor
Design Assistant — Tailor-made for modern hardware
Offload Advisor
estimates performance offloaded to accelerators
Roofline analysis
Optimizes CPU/GPU code for memory and computation
Vectorization Advisor
adds and optimizes vectorization
Threading Advisor
adds efficient threading functionality to non-threaded applications
Flow graph analyzer
Efficient creation and analysis flow graph