Tips and Tricks for Migrating CUDA* to SYCL*

The initial porting of your application from CUDA* to SYCL* using the Intel® DPC++ Compatibility Tool, or SYCLomatic, is quick and straightforward. However, there can be CUDA-specific routines in the original codebase that require extra attention.

This session looks at the details of the following advanced topics, some or all of which may be necessary to achieve successful CUDA-to-SYCL migration and performance portability for your target GPUs or AI accelerators:

• CUDA performance library mapping to specific oneAPI specification SYCL API libraries
• CPU offload kernel problem sizing
• CPU and GPU data caching, data transfer, and memory use optimization
• Unified Shared Memory versus Buffered Memory
• SYCL interoperability
• Resolving the use of custom Parallel Thread Execution (PTX) instructions
• Coexistence with MPI and distributed computing in HPC

Note It is recommended that you have a working knowledge of C and C++.

Skill level: All

This session features the following tools, which you can get as stand-alone versions or as part of the Intel® oneAPI Base Toolkit.

Note A stand-alone Intel DPC++ Compatibility Tool is not available.

• SYCLomatic
• Intel® oneAPI DPC++ Compiler
• Intel® Advisor
• Intel® VTune™ Profiler
• Intel® oneAPI Math Kernel Library (oneMKL)

• Get Started: Migrate CUDA to SYCL

See All Code Samples