Demonstrate maximal "double sqrt" GFLOPS performance for AVX2 CPUs

AVX2.vsqrtpd.cpp

/*
f=AVX2.vsqrtpd
g++ -O3 -fopenmp -Wall -Wextra -pedantic $f.cpp -o $f
cpplint --filter=-legal/copyright $f.cpp
cppcheck --enable=all --suppress=missingIncludeSystem $f.cpp --check-config

echo off | sudo tee /sys/devices/system/cpu/smt/control
echo 0 | sudo tee /proc/sys/kernel/perf_event_paranoid
perf stat -a -e cycles,instructions,task-clock ./$f

Output:
hermann@7950x:~$ ./$f
Starting hardware-bound benchmark using 16 threads...
... [AVX512F] vsqrtpd(mm512d,mm512d) completed
https://www.officedaytime.com/simd512e/simdimg/unop_qword_3.png
-------------------------------------------
Execution Time: 4.75232 seconds
Counter:        25,600,000,000
Total Compute:  204.8 double sqrt GFLOPS (counter * 8)
Performance:    43.0947 GFLOPS
-------------------------------------------
hermann@7950x:~$ 
*/
#include <omp.h>
#include <inttypes.h>
#include <iostream>
#include <chrono>   // NOLINT [build/c++11]

int main(int, char**) {
  const int iterations = 200000000;  // 2*10^8

  std::cout << "Starting hardware-bound benchmark using "
            << omp_get_max_threads() << " threads...\n";

  auto start_time = std::chrono::high_resolution_clock::now();

  #pragma omp parallel
  {
    for (int i = 0; i < iterations; ++i) {
asm __volatile__ (
            "vsqrtpd %%ymm0, %%ymm0 \n\t"
            "vsqrtpd %%ymm1, %%ymm1 \n\t"
            "vsqrtpd %%ymm2, %%ymm2 \n\t"
            "vsqrtpd %%ymm3, %%ymm3 \n\t"
            "vsqrtpd %%ymm4, %%ymm4 \n\t"
            "vsqrtpd %%ymm5, %%ymm5 \n\t"
            "vsqrtpd %%ymm6, %%ymm6 \n\t"
            "vsqrtpd %%ymm7, %%ymm7 \n\t"
            : // No outputs
            : // No inputs
            : "ymm0", "ymm1", "ymm2", "ymm3", "ymm4", "ymm5", "ymm6", "ymm7"
        );
    }
  }

  std::cout << "... [AVX256] vsqrtpd(mm256d,mm256d) completed\n";

  std::chrono::duration<double> duration =
    std::chrono::high_resolution_clock::now() - start_time;

  int64_t ops_per_loop = 4 * 8 * omp_get_max_threads();
  int64_t total_ops = ops_per_loop * iterations;
  int64_t giga_cnt = total_ops / 4;
  double giga_ops = total_ops / 1e9;
  double performance = giga_ops / duration.count();

  std::cout << "-------------------------------------------\n";
  std::cout << "Execution Time: " << duration.count() << " seconds\n";
  std::cout.imbue(std::locale(""));
  std::cout << "Counter:        " << giga_cnt << "\n";
  std::cout << "Total Compute:  " << giga_ops
            << " double sqrt GFLOPS (counter * 4)\n";
  std::cout << "Performance:    " << performance << " GFLOPS\n";
  std::cout << "-------------------------------------------\n";

  return 0;
}

On 8-socket 192C E7-8890v4 CPUs server with SMT disabled:

hermann@x3950-X6:~$ nproc
192
hermann@x3950-X6:~$ ./$f
Starting hardware-bound benchmark using 192 threads...
... [AVX256] vsqrtpd(mm256d,mm256d) completed
-------------------------------------------
Execution Time: 12.0587 seconds
Counter:        307,200,000,000
Total Compute:  1,228.8 double sqrt GFLOPS (counter * 4)
Performance:    101.902 GFLOPS
-------------------------------------------
hermann@x3950-X6:~$