openaccusergroup / openacc_concept_strategies_book Goto Github PK

Chapter 2, Mandelbrot example:

Accelerator version using PGI compiler gives a different result than the same code using gcc8 (for both openmp and nvptx+acc), pgi using openmp and the serial pgi, see results below:
tests> ./area.gcc.acc
Area of Mandlebrot set = 1.50787863 +/- 0.00037697
Time taken for calculation: 35.649632
tests> ./area.gcc.omp
Area of Mandlebrot set = 1.50787371 +/- 0.00037697
Time taken for calculation: 37.912720 on 8 threads
tests> ./area.pgi.omp
Area of Mandlebrot set = 1.50787863 +/- 0.00037697
Time taken for calculation: 27.200890 on 8 threads
tests> ./area.pgi.acc
Area of Mandlebrot set = 2.98259191 +/- 0.00074565
Time taken for calculation: 15.394787
tests> ./area.pgi.ser
Area of Mandlebrot set = 1.50787863 +/- 0.00037697
Time taken for calculation: 81.918041
tests>

Moreover, the pgi+acc version results are not stable, upon repeating the execution several times, different results were obtained:

tests> ./area.pgi.acc
Area of Mandlebrot set = 5.19458836 +/- 0.00129865
Time taken for calculation: 15.496788
tests> ./area.pgi.acc
Area of Mandlebrot set = 1.87656363 +/- 0.00046914
Time taken for calculation: 15.315470
tests> ./area.pgi.acc
Area of Mandlebrot set = 3.35120801 +/- 0.00083780
Time taken for calculation: 15.405854
tests>

The tests took place on an Asus laptop with 16GB RAM, i7 CPU, nvidia GeForce GTX-1050 GPU and Linux Mint.

Regards,

Guy Tel-Zur