The Past, Present, and Future of OpenACC
1 like•1,223 views
The document discusses the significance of compiler directives, specifically focusing on OpenACC and OpenMP, highlighting their roles in parallel programming and performance portability. It outlines the benefits of using compiler directives for rapid development, including maintaining single source code and ease of use. The status of OpenACC, its implementations, and the evolution towards improved support for complex data structures are also addressed.
![13
SOLVING DEEP COPY
Inform the compiler of the
shape of the data behind the
pointer.
struct T1 {
int N;
float* A;
#pragma acc policy(“shape")
shape( A[0:N] )
};
Develop policies for how the
data should be relocated.
struct T1 {
int N;
float* A;
#pragma acc policy(“shape")
shape( A[0:N] )
#pragma acc policy("boundary")
update(A[0:1],A[N-1:1)
};
Self-describing Structures Data Policies
*Syntax and functionality subject to change](https://image.slidesharecdn.com/jefflarkin-150301131425-conversion-gate01/85/The-Past-Present-and-Future-of-OpenACC-13-320.jpg)
The Past, Present, and Future of OpenACC
- 1. Jeff Larkin, NVIDIA OPENACC FEATURES FUTURE DIRECTIONS
- 2. 2 AGENDA • Why Compiler Directives? • OpenACC and OpenMP Status and Roadmap
- 4. 4 WHAT ARE COMPILER DIRECTIVES? program myscience ... serial code ... do k = 1,n1 do i = 1,n2 ... enddo enddo ... end program myscience CPU GPU Your original Fortran, C, or C++ code Insert portable compiler directives Compiler parallelizes code and manages data movement Programmer optimizes incrementally Designed for multi-core CPUs, GPUs & many-core Accelerators program myscience ... serial code ... !$acc parallel loop do k = 1,n1 do i = 1,n2 ... enddo enddo ... end program myscience
- 5. 5 WHY USE COMPILER DIRECTIVES? Single Source Code No need to maintain multiple code paths High Level Abstract away device details, focus on expressing the parallelism and data locality Low Learning Curve Programmer remains in same language and adds directives to existing code Rapid Development Fewer code changes means faster development
- 6. 6 1,000,000’s Early Adopters Research Universities Supercomputing Centers Oil & Gas WHY SUPPORT COMPILER DIRECTIVES? 100,000’s 2004 Present Increased Usability, Broader Adoption CAE CFD Finance Rendering Data Analytics Life Sciences Defense Weather Climate Plasma Physics
- 7. 7 3 WAYS TO ACCELERATE APPLICATIONS Applications Libraries “Drop-in” Acceleration Programming Languages Compiler Directives Maximum Flexibility Easily Accelerate Applications
- 8. 8 OpenACC & OpenMP Status
- 9. 9 COMPILER DIRECTIVES TODAY OpenACC Focused solely on compiler directives for accelerators Quickly moving Performance Portability a primary consideration Descriptive approach to parallel programming OpenMP Addresses a broad range of parallel programming challenges More measured approach Less focus on performance portability Prescriptive approach to parallel programming Choice of two great options
- 10. 10 OPENACC IMPLEMENTATIONS OpenACC 2.0 launched December 2013 OpenACC 2.0 launched January 2014 OpenACC 2.0 Targeted for Late 2014 Tremendous interest in academia: • accULL – U. La Laguna/EPCC • Omni – U. of Tsukuba • OpenARC – Oak Ridge NL • OpenUH – U. of Houston • RoseACC – U of DE & LLNL Compilers Debuggers and Profilers OpenACC 2.0 Targeted for 2015
- 11. 11 11 OPENACC 2.NEXT Deep-Copy Support for more complex data structures Improves C++ class support Tools API Standard API for profiling tools Default to present_or behavior for data clauses Tell us what else you need [email protected]
- 12. 12 WHAT IS DEEP COPY? On Host Shallow Copy Deep Copy Deep copy is required for structures with pointer members to work properly, but also useful where only part of a structure is needed on the device. • Users can do this today manually, but we want it to be automatic
- 13. 13 SOLVING DEEP COPY Inform the compiler of the shape of the data behind the pointer. struct T1 { int N; float* A; #pragma acc policy(“shape") shape( A[0:N] ) }; Develop policies for how the data should be relocated. struct T1 { int N; float* A; #pragma acc policy(“shape") shape( A[0:N] ) #pragma acc policy("boundary") update(A[0:1],A[N-1:1) }; Self-describing Structures Data Policies *Syntax and functionality subject to change
- 14. 14 OPENACC & UNIFIED MEMORY With Unified Memory (or equivalent) shallow copying no longer results in runtime error, so does deep-copy become irrelevant? Dereferencing the copied pointer no-longer results in error. Hardware support for Unified Memory would allow for data to migrate on access. Deep Copy support in OpenACC is still beneficial because Data can be moved before it’s needed without penalty Code will be portable to devices without Unified Memory support
- 15. 17 17 THE “MERGE” Both specifications bring value to the market OpenACC – more focused, more agile, descriptive approach OpenMP – more cautious, more mature overall, prescriptive approach The specifications are constantly converging, even if they never eventually merge. Use the best tool available to you today, the restructuring and knowledge from each are portable. In other words: the syntax isn’t the hard part. Does having 2 specifications cause confusion?
- 16. 19 SHOULD I JUST WAIT FOR OPENMP4 SUPPORT? No!
- 17. THANK YOU