Code

Usage

• clang -fopenmp -L/usr/local/opt/llvm/lib -I/usr/local/opt/llvm/include <filename> && ./a.out
• Works with clang version 9.0.1 , x86_64-apple-darwin19.3.0

Code Examples

1. Hello World
2. Pi
3. Mandelbrot Set

• 1_hello_world.c

  • hello world of OpenMP
  • code in structured block following "# pragma omp parallel" is executed by every thread possibly available

  # pragma omp parallel
  

• 2_serial_pi.c

  • serial program to calculate pi via integration
  • pragma omp parallel - get me default number of system threads
  • my system has 12 threads. 2 threads per core

• 2_parallel_pi_v1.c

  • Extends v1. Calculate pi with OpenMP
  • Key ideas -
    • Promote scalar (sum) to an array dimensioned by number of threads to avoid race condition
    • Only one thread should copy the number of threads to the global value to make sure multiple threads writing to the same address don’t conflict.
    • For loop uses the most common trick in SPMD programs to create a cyclic/round robin distribution of loop(threads) iterations
  • run with 1,2,4,6 threads. scalability of threads is terrible. because of false sharing
  • worse performance than v1

• 2_parallel_pi_v2.c

  • use padding to prevent false sharing. forces data of each thread to be on distinct cache lines
  • better performance than v1
  • but hacky poor solution. how would you know the size of L1 cache line? wouldn't work across architectures

• 2_parallel_pi_v3.c

  • Use synchronization to prevent false sharing and also prevent use of padding
  • Create a scalar local to each thread to accumulate partial sums.
    • Now sum variable on stack of each loop. False sharing automatically eliminated
  • Pi summation within the critical section now, since the local variable sum will be lost with stack once threads end
  • Performance as good as v2. Portable code.
  • Lesson: Put #pragma omp critical on the code where work done is minimal. Do not serialize your program. (Slide 71)

• 2_parallel_pi_v4.c

  • Start with serial pi program and add "omp for","reduction"
  • As good performance as it gets

• 2_parallel_pi_v4_2.c

  • Better implementation of 2_parallel_pi_v4.c
  • Just 2 line change wrt serial version
    • Goal for parallel programming. Serial code doesn't break.

• 3_mandel_bug.c

  • Buggy version of Mandelbrot set. Incorrect answer

• 3_mandel_fix.c. Fixed

  1. firstprivate(eps) private(c, i, j)
  2. void testpoint(struct d_complex c) and void testpoint(struct d_complex);
  3. #pragma omp atomic
        numoutside++;
  4. testpoint(c);