Skip to content

Latest commit

 

History

History
87 lines (53 loc) · 3.44 KB

README.md

File metadata and controls

87 lines (53 loc) · 3.44 KB

SYCL compiler for dynamic rebinding

This project is for providing transparency and dynamic rebinding of the use of heterogeneous accelerators on edge micro data centers

Ready for dynamic rebinding

Compile sycl code for multiple target backends with linker option to specify file offset of clang offload section

e.g. clang++ -fsycl -fsycl-targets=[backend list] main.cpp -Xlink [linker arg] (-Xlink [linker arg]) -o out.exe

[backends list] : comma seperated list

spir64_x86_64 for Intel CPU backend  
spir64_gen for Intel GPU (ComputeCpp) backend  
spir64_fpga for Intel FPGA backend(emulation or real device)  
nvptx64 for NVIDIA PTX backend* (llvm should be configured *cuda enabled and compiled to use this target)

[linker arg] : argument of GNU linker 'ld'

--section-start <section_name>=<address>  
<section_name> : __CLANG_OFFLOAD_BUNDLE__sycl-{target backend}    
[address] should be greater than any other section's address

Docker image

sunginh/emdc-oneapi-base-cuda:v1.0 - for sycl supports Intel CPU, NVIDIA GPU, INTEL FPGA PAC D5005

(test)sunginh/emdc-ow-cuda-11.8.0:v0.2 - for baseline

(test)sunginh/emdc-ow-cuda-11.8.0:v0.83 - for persistent thread model

Object Storage

This project adopts MinIO as Object Storage which is compatible with AWS S3.

https://min.io/

https://github.com/minio/minio

Work-in-progress

Reduce startup latency for NVIDIA GPU: context creation time(~220ms)

Reduce startup latency for Intel FPGA: Program FPGA bit stream(~2s)

Release dense funtion bottleneck of GPU driver: NVIDIA GPU driver hold two types of lock, RMAPI lock for blocking APIs and GPU locks for non-blocking APIs.

Forked from

Intel LLVM-based projects (https://github.com/intel/llvm)

README.md from Intel LLVM-based projects as below:

oneAPI Data Parallel C++ compiler

See sycl branch and DPC++ Documentation.

Linux Post Commit Checks Generate Doxygen documentation

DPC++ is an open, cross-architecture language built upon the ISO C++ and Khronos SYCL* standards. DPC++ extends these standards with a number of extensions, which can be found in sycl/doc/extensions directory.

Late-outline OpenMP* and OpenMP* Offload

See openmp branch.

License

See LICENSE.txt for details.

Contributing

See CONTRIBUTING.md for details.

*Other names and brands may be claimed as the property of others.

Related Accomplishments

국내 특허 출원 1건 완료

  • 멀티테넌트 GPU 환경에서 메모리 자원을 고려한 선제적 동시-스케줄링

국내 특허 등록 2건 진행중

  • 이종 컴퓨팅 클라우드의 컴퓨팅 유닛 동적 재조립 방법 및 장치
  • 서비스형 함수 환경에서의 함수 실행 방법

해외 특허 등록 1건 진행중

  • Method of rebinding computing unit in heterogeneous computing clouds and apparatus thereof