Skip to main content

Source repo: sdsc-summer-institute-2021 | Branch: main | Last synced: 2026-04-24 10:27:17.425 UTC

OpenACC samples

SDSC Summer Institute

August 5, 2021

Material by Andreas Goetz (agoetz@sdsc.edu). Partly based on work by Mark Harris - NVIDIA corporation.

Directory content as explained below

saxpy (Single Precision A * X Plus Y)​

See directoy saxpy/.

saxpy.c C program that performs single precision vector addition (y = a*x + y). Can be compiled with standard C compiler or PGI pgcc compiler with accelerator directives.

pgcc saxpy.c -o saxpy-cpu.x
pgcc saxpy.c -acc -Minfo=accel -o saxpy-gpu.x

Jacobi solver for 2D Laplace equation​

See directory laplace-2d/.

jacobi.c and jacobi.f90 C and Fortran versions of a Jacobi solver for 2D Laplace equation.

Compile and benchmark as follows:

Benchmarks on SDSC Comet node, using K80 GPU​

First make sure to have the PGI compiler loaded:

module load pgi

Serial version:

pgf90 jacobi.f90 -fast -o jacobi-pgf90.x
./jacobi-pgf90.x      # completes in 50 seconds

pgcc jacobi.c -fast -o jacobi-pgcc.x
./jacobi-pgcc.x       # completes in 71 seconds

OpenMP parallel version:

pgf90 jacobi-omp.f90 -fast -mp -Minfo=mp -o jacobi-pgf90-omp.x
export OMP_NUM_THREADS=2
./jacobi-pgf90-omp.x   # completes in 33 seconds
export OMP_NUM_THREADS=4
./jacobi-pgf90-omp.x   # completes in 25 seconds
export OMP_NUM_THREADS=6
./jacobi-pgf90-omp.x   # completes in 24 seconds
export OMP_NUM_THREADS=12
./jacobi-pgf90-omp.x   # completes in 23 seconds
export OMP_NUM_THREADS=24
./jacobi-pgf90-omp.x   # completes in 22 seconds

pgcc jacobi-omp.c -fast -mp -Minfo=mp -o jacobi-pgfcc-omp.x
export OMP_NUM_THREADS=2
./jacobi-pgcc-omp.x    # completes in 41 seconds
export OMP_NUM_THREADS=4
./jacobi-pgcc-omp.x    # completes in 26 seconds
export OMP_NUM_THREADS=6
./jacobi-pgcc-omp.x    # completes in 24 seconds
export OMP_NUM_THREADS=12
./jacobi-pgcc-omp.x    # completes in 23 seconds
export OMP_NUM_THREADS=24
./jacobi-pgcc-omp.x    # completes in 23 seconds

OpenACC GPU accelerated version:

pgf90 jacobi-acc.f90 -acc -Minfo=accel -o jacobi-pgf90-acc.x
./jacobi-pgf90-acc.x   # completes in 5 seconds

pgcc jacobi-acc.c -acc -Minfo=accel -o jacobi-pgcc-acc.x
./jacobi-pgcc-acc.x   # completes in 6 seconds

Benchmarks on V100 test node​

For comparison, see timings on a system with following setup:

  • 2 x 16-core Intel Xeon Gold 6130 CPU @ 2.10 GHz
  • Nvidia V100-SXM2 GPU

Serial version:

module load pgi
pgicc --version
pgcc 18.10-1 64-bit target on x86-64 Linux -tp skylake 

pgf90 jacobi.f90 -fast -o jacobi-pgf90.x
./jacobi-pgf90.x      # completes in 44.3 seconds

pgcc jacobi.c -fast -o jacobi-pgcc.x
./jacobi-pgcc.x       # completes in 48.2 seconds

OpenMP parallel version:

pgf90 jacobi-omp.f90 -fast -mp -Minfo=mp -o jacobi-pgf90-omp.x
export OMP_NUM_THREADS=2
./jacobi-pgf90-omp.x   # completes in 24 seconds
export OMP_NUM_THREADS=4
./jacobi-pgf90-omp.x   # completes in 13 seconds
export OMP_NUM_THREADS=8
./jacobi-pgf90-omp.x   # completes in 8 seconds
export OMP_NUM_THREADS=16
./jacobi-pgf90-omp.x   # completes in 6 seconds
export OMP_NUM_THREADS=32
./jacobi-pgf90-omp.x   # completes in 6 seconds

pgcc jacobi-omp.c -fast -mp -Minfo=mp -o jacobi-pgfcc-omp.x
export OMP_NUM_THREADS=2
./jacobi-pgcc-omp.x    # completes in 25 seconds
export OMP_NUM_THREADS=4
./jacobi-pgcc-omp.x    # completes in 13 seconds
export OMP_NUM_THREADS=6
./jacobi-pgcc-omp.x    # completes in 8 seconds
export OMP_NUM_THREADS=12
./jacobi-pgcc-omp.x    # completes in 6 seconds
export OMP_NUM_THREADS=24
./jacobi-pgcc-omp.x    # completes in 6 seconds

OpenACC GPU accelerated version:

pgf90 jacobi-acc.f90 -acc -Minfo=accel -o jacobi-pgf90-acc.x
./jacobi-pgf90-acc.x   # completes in 1.3 seconds

pgcc jacobi-acc.c -acc -Minfo=accel -o jacobi-pgcc-acc.x
./jacobi-pgcc-acc.x   # completes in 1.3 seconds