Awesome

Introduction

The directories in this repository contain code examples for the course of OpenMP GPU-offloading at Paderborn Center for Parallel Computing (PC²), Paderborn University. The sub-directories are generally organized as:

• src: source code
• docs: documentation
• tests: some tests

Some highlights of the codes in this repository:

• The performance of our saxpy implemented by using OpenMP GPU-offloading is as good as cublasSaxpy in CUBLAS. See case 7 in 05_saxpy/src/asaxpy.c for details.

• The GPU shared memory has not been standardized in OpenMP API Specification (Version 5.0 Nov. 2018). To optimize the performance of matrix multiplication by using OpenMP GPU-offloading, i) case 6 in 10_matMul/src/matMulAB.c implements a register blocking algorithm and ii) case 8 in the same source code file implements a common GPU-based tiled algorithm by blocking the local shared memory in a very tricky manner and the OpenMP code resembles CUDA.

List of Projects

• 00_build_OpenMP_offload

  Documentation and scripts for building GCC as well as Clang/LLVM with OpenMP support for Nvidia GPU offloading.

• 01_accelQuery

  accelQuery searches accelerator(s) on a heterogeneous computer. Accelerator(s), if found, will be enumerated with some basic info.

• 02_dataTransRate

  dataTransRate gives the data transfer rate (in MB/sec) from src to dst.

  The possible situations are:

  • h2h: src = host and dst = host
  • h2a: src = host and dst = accel
  • a2a: src = accel and dst = accel

  NOTE:

  • A bug in Clang 9.0.1 has been fixed in Clang 11.
  • The data transfer rata for a2a is still lower than our expectation.

• 03_taskwait

  taskwait checks the taskwait construct for the deferred target task.

  NOTE:

  • Asynchronous offloading hasn't been implemented in the GCC 9.2 compiler.
  • Asynchronous offloading is available in Clang 11.

• 04_scalarAddition

  scalarAddition adds two integers on host and accelerator, and also compares the performance.

• 05_saxpy

  saxpy performs the saxpy operation on host as well as accelerator. The performance (in MB/s) for different implementations is also compared.

• 08_distThreads

  distThreads demonstrates the organization of threads and teams in a league on GPU.

• 09_matAdd

  matAdd performs matrix addition (A +=B) in single-precision on GPU. The performance (in GB/s) for different implementations is compared and the numerical results are also verified.

• 10_matMul

  matMul performs matrix multiplication in single-precision on GPU. The performance (in GFLOPS) for different implementations is compared and the numerical results are also verified.